Skip to content

Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.

Preview first, prompt box improvements and delete in Your creations; install into folders with spaces (0.3.2) (#86)

Bowen Xue635423b6517f
diff --git a/README.md b/README.md
index 04501c3..71c99b8 100644
--- a/README.md
+++ b/README.md
@@ -19,7 +19,8 @@ https://github.com/user-attachments/assets/ecda7d0d-7fbe-4e0c-8c29-8f3315bafc15
## News
-- **2026-10-07** · **[v0.3.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.0): a faster int8 model and prompt enhancement.** GeForce RTX 40/50 and RTX 30-series cards switch to an int8 model that is faster and closer to the original quality; an RTX 3090 makes a 5-second video about 2.3x faster. Existing installations upgrade in one click, and an optional local prompt enhancer rewrites prompts for MiniMax H3.
+- **2026-10-07** · **[v0.3.2](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.2): preview first.** See a half-resolution preview first and finish only the videos you like.
+- **2026-10-07** · **[v0.3.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.0): a faster int8 model.** GeForce and RTX 30-series cards switch to int8; an RTX 3090 makes a 5-second video about 2.3x faster.
- **2026-10-06** · **[Gallery](https://freevideo-community.pages.dev/#gallery) is live.** Watch 20-second clips made with FreeVideo, and the four quality levels side by side.
- **2026-10-06** · **[v0.2.3](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.2.3): videos carry their workflow.** Drop a FreeVideo video onto the ComfyUI canvas to restore its prompt, seed and settings. Earlier videos can get theirs too.
- **2026-10-05** · **[v0.2.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.2.0): four quality levels.** Choose Light, Medium, High or Max for each video; higher levels give higher quality but take longer. Results can be exported as sharing images or videos with the generation time and GPU.
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 5b7c285..c79b817 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -19,7 +19,8 @@ https://github.com/user-attachments/assets/ecda7d0d-7fbe-4e0c-8c29-8f3315bafc15
## 更新动态
-- **2026-10-07** · **[v0.3.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.0):更快的 int8 模型和提示词增强。** GeForce RTX 40/50 系和 RTX 30 系显卡改用更快、画质更接近原版的 int8 模型,RTX 3090 生成 5 秒视频快约 2.3 倍;已安装的用户可一键升级。新增可选的本地提示词增强,把提示词改写成适合 MiniMax H3 的写法。
+- **2026-10-07** · **[v0.3.2](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.2):先出预览。** 先看半分辨率预览,满意的再继续二采。
+- **2026-10-07** · **[v0.3.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.0):更快的 int8 模型。** GeForce 和 RTX 30 系显卡改用 int8,RTX 3090 生成 5 秒视频快约 2.3 倍。
- **2026-10-06** · **[作品展示](https://freevideo-community.pages.dev/#gallery)上线。** 观看 FreeVideo 生成的 20 秒视频,以及四档质量的并排对比。
- **2026-10-06** · **[v0.2.3](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.2.3):视频自带工作流。** 把 FreeVideo 生成的视频拖到 ComfyUI 画布上,即可还原提示词、种子和全部参数,之前的视频也能补上。
- **2026-10-05** · **[v0.2.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.2.0):四档质量。** 每次生成可选择轻量、标准、精细或极致,档位越高,生成质量越高,但耗时更长;生成结果可导出为带生成耗时和显卡信息的分享图片或视频。
diff --git a/freevideo_engine/adaptive.py b/freevideo_engine/adaptive.py
index d4f8b88..215c542 100644
--- a/freevideo_engine/adaptive.py
+++ b/freevideo_engine/adaptive.py
@@ -621,6 +621,11 @@ def complete_observation(metrics, canvas, output, artifacts, telemetry):
from .system import memory_complete, memory_peak
from .validation import _retained_array_header, finite_number
expected = canvas['frames']
+ preview = metrics.get('preview') is True
+ if preview:
+ # A preview decodes its first pass at the planned smaller canvas.
+ first = (metrics.get('sampling_plan') or {}).get('first') or {}
+ canvas = dict(canvas, width=first.get('width'), height=first.get('height'))
if (metrics.get('success') is not True or metrics.get('phase') != 'complete'
or metrics.get('finite_latents') is not True or metrics.get('frames') != expected
or any(metrics.get('geometry', {}).get(k) != canvas[k] for k in ('width','height','frames','fps'))):
@@ -629,6 +634,8 @@ def complete_observation(metrics, canvas, output, artifacts, telemetry):
times = metrics.get('step_seconds', [])
from .two_pass import steps as sampling_steps, same_strategy
total_steps = sampling_steps(metrics.get('sampling_plan'))
+ if preview:
+ total_steps = metrics['sampling_plan']['base_steps']
if (type(steps) is not int or steps != canvas.get('steps', 8) or len(times) != total_steps
or not same_strategy(metrics, canvas) or not all(finite_number(t) and t > 0 for t in times)):
raise ValueError('A probe or incomplete sample cannot become a resource observation.')
@@ -653,6 +660,10 @@ def complete_observation(metrics, canvas, output, artifacts, telemetry):
# retry. Do not teach placement/ETA that a full request cost this little
# time or occupied only the decoder's memory.
return None
+ if preview or metrics.get('first_pass_source') == 'preview':
+ # A preview stops after its first pass, and its upscale reuses that pass
+ # from an earlier request. Neither costs what a whole request costs.
+ return None
ram, gpu = telemetry.get('ram', {}), telemetry.get('gpu', {})
counts = (ram.get('ram_observation_samples'), gpu.get('samples'))
if (not memory_complete(ram) or any(type(n) is not int or n <= 0 for n in counts)
diff --git a/freevideo_engine/bootstrap.py b/freevideo_engine/bootstrap.py
index 4fa40ab..19d4ac0 100644
--- a/freevideo_engine/bootstrap.py
+++ b/freevideo_engine/bootstrap.py
@@ -26,7 +26,8 @@ from .monitoring import save
from .install_tuning import build_parallelism, required_models, cache_compatible, wheel_key
from .terminal_ui import TerminalUI, LogProgress
from .locking import runtime_lock, LOCK_ENV
-from .environments import ENVIRONMENTS, environment_names, select_layout, role_pythons, constraints_file, bootstrap_versions
+from .environments import (ENVIRONMENTS, environment_names, select_layout, role_pythons, constraints_file,
+ bootstrap_versions, uv_file_arguments)
from . import network
from . import processes
from .system import install_root, venv_python, system_memory, nvidia_smi, memory_sample, curl_executable, missing_curl_message
@@ -1032,10 +1033,11 @@ class Installer:
from .package_progress import PackageOutput
if self.cancel.is_set():
raise RuntimeError(self.state.get('resource_guard', 'Installation cancelled.'))
+ args, cwd = uv_file_arguments(args, cwd)
with self.state_lock:
index = len(self.state['steps'])
log = self.run_dir / ('%02d-%s.log' % (index, label))
- row = {'label': label, 'command': list(map(str, args)), 'log': str(log), 'status': 'running'}
+ row = {'label': label, 'command': args, 'log': str(log), 'status': 'running'}
self.state['steps'].append(row)
save(self.run_dir / 'status.json', self.state)
key = str(index) + '-' + label
diff --git a/freevideo_engine/cli.py b/freevideo_engine/cli.py
index ce45838..78a3777 100644
--- a/freevideo_engine/cli.py
+++ b/freevideo_engine/cli.py
@@ -103,6 +103,13 @@ def main():
help='First-pass steps (default: 8, or explicit profile). Changes may reduce quality.')
generate.add_argument('--refine-steps', type=int, choices=range(1, 32), default=3, metavar='1..31',
help='Second-pass steps (default: 3, independent schedule). Other counts use the original tail.')
+ upscale = generate.add_mutually_exclusive_group()
+ upscale.add_argument('--preview', action='store_true',
+ help='Two-pass only: stop after the half-resolution first pass, save it as a quick preview, '
+ 'and keep that pass so the preview can be upscaled later')
+ upscale.add_argument('--refine-from', type=Path, metavar='PREVIEW.mp4',
+ help="Upscale a finished preview: reuse its first pass, then the latent upscale, the refinement "
+ "steps and a full-resolution decode. Use the preview's own prompt, seed, size and inputs")
generate.add_argument('--out', type=Path, required=True)
generate.add_argument('--width', type=int, default=1344)
generate.add_argument('--height', type=int, default=768)
diff --git a/freevideo_engine/comfy_assets.py b/freevideo_engine/comfy_assets.py
index 0badb58..6609856 100644
--- a/freevideo_engine/comfy_assets.py
+++ b/freevideo_engine/comfy_assets.py
@@ -153,9 +153,12 @@ def output_summary(report, relative_video):
geometry=report.get('geometry', {}),
sampling_plan=report.get('sampling_plan'),
reference_trims=report.get('encoding', {}).get('reference_trims') or [],
+ preview=bool(report.get('preview')), upscaled_preview=bool(report.get('upscaled_preview')),
video=relative_video.as_posix(), report=relative_video.with_suffix(
'.debug.json' if report.get('diagnostic_file') == relative_video.with_suffix('.debug.json').name
else '.request.json').as_posix())
+ if report.get('preview'):
+ result['preview_geometry'] = report['preview'].get('geometry')
if report.get('device_backend') == 'mps':
policy = report.get('profile', {}).get('policy', {})
hardware = report.get('runtime_hardware', {})
diff --git a/freevideo_engine/comfy_bridge.py b/freevideo_engine/comfy_bridge.py
index 0e76536..07a3072 100644
--- a/freevideo_engine/comfy_bridge.py
+++ b/freevideo_engine/comfy_bridge.py
@@ -366,6 +366,9 @@ def progress_message(event):
return dict(label=label, timing_phase='encoding') if label else None
if name == 'sampling_plan':
first, second = event.get('first', {}), event.get('second')
+ if event.get('enabled') and event.get('preview'):
+ return dict(label='First-pass preview · %d × %d · %d steps' % (first['width'], first['height'], event.get('base_steps', 8)),
+ detail='Run the second pass from the result', sampling_plan=event)
if event.get('enabled') and second:
return dict(label='Two-pass · %d × %d → %d × %d · %d + %d steps' %
(first['width'], first['height'], second['width'], second['height'],
@@ -374,6 +377,15 @@ def progress_message(event):
return dict(label='Single-pass · %d steps' % event.get('base_steps', 8), detail=event.get('reason'), sampling_plan=event)
if name == 'latent_upscaler_prepare':
return dict(label='Preparing two-pass upscaler', timing_phase='load')
+ if name == 'first_pass_reused':
+ done, total = event.get('completed_steps', 8), event.get('total', 11)
+ return dict(label='Continuing from the preview' if event.get('source') == 'preview'
+ else 'Reusing the completed first pass', phase='sampling', stage='first_pass_reused',
+ done=done, total=total, display_fraction=done / total, timing_phase='sampling',
+ estimated_step_seconds=None, step_elapsed_seconds=0., remaining_seconds=None,
+ uniform_remaining_steps=False)
+ if name == 'preview_mismatch':
+ return dict(label='The preview no longer matches; sampling its first pass again', stage='preview_mismatch')
if name == 'latent_upscale':
done, total = event.get('completed_steps', 8), event.get('total', 10)
return dict(label='Upscaling before the second pass', phase='sampling', stage='latent_upscale',
@@ -455,6 +467,70 @@ def progress_message(event):
return None
+def preview_request(output_directory, relative):
+ """The saved request of a finished preview in this output folder."""
+ if not isinstance(relative, str) or not relative.strip():
+ raise ValueError('Choose a preview to upscale')
+ output_root = Path(output_directory).resolve()
+ video = (output_root / relative).resolve()
+ try:
+ video.relative_to(output_root / 'FreeVideo')
+ except ValueError:
+ raise ValueError('Choose a FreeVideo preview from this output folder') from None
+ path = video.parent / 'comfy-request.json'
+ state = json.loads(path.read_text(encoding='utf-8')) if path.is_file() else {}
+ if (not isinstance(state, dict) or state.get('preview') is not True or state.get('status') != 'complete'
+ or Path(state.get('output', '')).resolve() != video or not isinstance(state.get('command'), list)
+ or 'generate' not in state['command']):
+ raise ValueError('This video is not a finished FreeVideo preview')
+ if not video.is_file():
+ raise ValueError('The preview video is no longer there; generate it again')
+ return video, state
+
+
+def upscale_command(state, output, video, resources):
+ """The preview's own generate arguments, refining it into a new output."""
+ tail = state['command'][state['command'].index('generate') + 1:]
+ replaced = {'--' + field.replace('_', '-') for field in list(state.get('resources') or {}) + list(resources)}
+ arguments, index = [], 0
+ while index < len(tail):
+ flag = tail[index]
+ if flag == '--preview':
+ index += 1
+ continue
+ if flag in replaced or flag == '--out':
+ if flag == '--out':
+ arguments += ['--out', str(output)]
+ index += 2
+ continue
+ arguments.append(flag)
+ index += 1
+ for field, reserve in resources.items():
+ if reserve is not None:
+ arguments += ['--' + field.replace('_', '-'), str(reserve)]
+ return arguments + ['--refine-from', str(video)]
+
+
+def without_upscale(graph, node_id, refine_from):
+ import copy
+ graph = copy.deepcopy(graph)
+ inputs = ((graph.get('prompt') or {}).get(str(node_id)) or {}).get('inputs')
+ if isinstance(inputs, dict):
+ inputs.update(refine_from='', preview=False)
+ for node in (graph.get('workflow') or {}).get('nodes') or []:
+ values = node.get('widgets_values')
+ if str(node.get('id')) == str(node_id) and isinstance(values, list) and values and values[-1] == refine_from:
+ # Preview first and Upscale preview are the node's last two widgets.
+ values[-1] = ''
+ if len(values) > 1 and type(values[-2]) is bool:
+ values[-2] = False
+ return graph
+
+
+def argument(arguments, flag, default=None):
+ return arguments[arguments.index(flag) + 1] if flag in arguments else default
+
+
def engine_environment(root, source, environ=None):
if sys.platform == 'darwin':
from .macos_bootstrap import environment
@@ -467,11 +543,29 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
source=None, environ=None, metadata=None, progress=None, interrupted=None,
release_models=None, export_inputs=None, two_pass=True, encoder_prewarm=None,
force_regenerate=False, base_steps=8, refine_steps=3, comfy_metadata=None,
- prompt_rewrite_report=None):
+ prompt_rewrite_report=None, preview=False, refine_from=None):
if type(two_pass) is not bool:
raise ValueError('Two-pass generation must be a boolean')
if type(force_regenerate) is not bool:
raise ValueError('Force regeneration must be a boolean')
+ if type(preview) is not bool:
+ raise ValueError('Preview must be a boolean')
+ refine_from = refine_from if isinstance(refine_from, str) and refine_from.strip() else None
+ upscale = None
+ if refine_from:
+ # An upscale continues the preview's own request: its prompt, seed,
+ # canvas, steps and inputs, not whatever the editor shows now.
+ upscale = preview_request(output_directory, refine_from)
+ _, previous = upscale
+ arguments = previous['command'][previous['command'].index('generate') + 1:]
+ prompt = (Path(previous['output']).parent / 'prompt.txt').read_text(encoding='utf-8')
+ width, height = previous['geometry']['width'], previous['geometry']['height']
+ seconds, seed = previous['geometry']['seconds'], previous['seed']
+ two_pass, preview = True, False
+ base_steps = int(argument(arguments, '--base-steps', 8))
+ refine_steps = int(argument(arguments, '--refine-steps', 3))
+ from .preview import validate_flags
+ validate_flags(preview, upscale, two_pass)
from .two_pass import validate_steps
validate_steps(base_steps, refine_steps, two_pass)
canvas = validate_request(prompt, width, height, seconds, seed)
@@ -500,11 +594,20 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
'--width', str(width), '--height', str(height), '--seconds', str(seconds), '--seed', str(seed),
'--two-pass' if two_pass else '--no-two-pass',
'--base-steps', str(base_steps), '--refine-steps', str(refine_steps)]
- for field, reserve in resources.items():
- if reserve is not None:
- command += ['--' + field.replace('_', '-'), str(reserve)]
+ if upscale:
+ command = command[:command.index('generate') + 1] + upscale_command(upscale[1], output, upscale[0], resources)
+ else:
+ for field, reserve in resources.items():
+ if reserve is not None:
+ command += ['--' + field.replace('_', '-'), str(reserve)]
+ if preview:
+ command.append('--preview')
state = {'status': 'starting', 'geometry': canvas, 'seed': seed, 'installation': str(root),
'command': command, 'output': str(output), 'source': str(source), 'resources': resources}
+ if preview:
+ state['preview'] = True
+ if upscale:
+ state['upscaled_preview'] = str(upscale[0])
from .diagnostic_resources import encoder_prewarm as prewarm_summary
state['encoder_prewarm'] = prewarm_summary(encoder_prewarm)
save(run / 'comfy-request.json', state)
@@ -534,7 +637,16 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
if interrupted:
interrupted()
extra = {}
- if export_inputs:
+ media_base = run
+ if upscale:
+ # The preview's retained inputs, exactly as its request used them.
+ arguments = command[command.index('generate') + 1:]
+ media_base = Path(upscale[1]['output']).parent
+ if '--media' in arguments:
+ extra['media'] = json.loads(Path(argument(arguments, '--media')).read_text(encoding='utf-8'))
+ if '--conditioning' in arguments:
+ extra['conditioning'] = argument(arguments, '--conditioning')
+ elif export_inputs:
send_progress({'label': 'Retaining and checking input media', 'timing_phase': 'encoding'})
extra = export_inputs(run, canvas) or {}
if extra:
@@ -549,11 +661,13 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
from .media_request import task_for
from .two_pass import plan
planned = plan(canvas, two_pass, task_for(extra.get('media', {})), base_steps=base_steps, refine_steps=refine_steps)
- identify = lambda check: request_key(prompt, seed, canvas, planned, extra, machine,
+ # A preview is its own result; an upscale is always run from its preview.
+ keyed = dict(planned, preview=True) if preview else planned
+ identify = lambda check: request_key(prompt, seed, canvas, keyed, extra, machine,
resources, source, environment, inspection=check)
state['result_cache'] = dict(enabled=False, hit=False, forced=force_regenerate)
reused = None
- if not force_regenerate:
+ if not force_regenerate and not upscale:
send_progress({'label': 'Checking saved video', 'timing_phase': 'encoding'})
def inspect_saved(check):
key = identify(check)
@@ -579,7 +693,7 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
preparation_started = time.monotonic()
try:
# Reference audio selects its own tables; match the encoder's choice.
- preparation = prepare_sampling_assets(root, machine, planned, engine_task(extra.get('media', {}), run),
+ preparation = prepare_sampling_assets(root, machine, planned, engine_task(extra.get('media', {}), media_base),
progress=asset_progress, interrupted=interrupted, environ=environment)
except BaseException:
elapsed = time.monotonic() - preparation_started
@@ -627,12 +741,18 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
completed_steps = steps(report.get('sampling_plan'))
expected_plan = plan(canvas, two_pass, report.get('profile', {}).get('engine', {}).get('task', 't2va'),
base_steps=base_steps, refine_steps=refine_steps)
+ shown = canvas
+ if preview:
+ # A preview decodes its first pass at the planned smaller canvas.
+ completed_steps = expected_plan['base_steps']
+ shown = dict(canvas, width=expected_plan['first']['width'], height=expected_plan['first']['height'])
if (report.get('success') is not True or engine.get('success') is not True
or not output.is_file() or not output.stat().st_size
or len(engine.get('step_seconds', [])) != completed_steps
or engine.get('sampling_plan') != report.get('sampling_plan')
or report.get('sampling_plan') != expected_plan
- or any(engine.get('geometry', {}).get(k) != canvas[k] for k in ('width', 'height', 'frames'))):
+ or bool(report.get('preview')) != preview or bool(report.get('upscaled_preview')) != bool(upscale)
+ or any(engine.get('geometry', {}).get(k) != shown[k] for k in ('width', 'height', 'frames'))):
from .failure_details import generation_failure
raise RuntimeError('FreeVideo did not complete the requested video.\n' + generation_failure(run))
if comfy_metadata:
@@ -651,7 +771,8 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
return check.measure('output', lambda: cache.remember(cache_key, output, inspection=check))
stored, timing = inspect_bounded(store_result, interrupted=interrupted)
state['result_cache'].update(stored=stored is True, store_inspection=timing)
- send_progress({'label': 'Video + audio saved', 'phase': 'complete', 'done': completed_steps, 'total': completed_steps})
+ send_progress({'label': 'Preview saved' if preview else 'Video + audio saved', 'phase': 'complete',
+ 'done': completed_steps, 'total': completed_steps})
return output
except BaseException as error:
cancelled = isinstance(error, KeyboardInterrupt) or type(error).__name__ in ('InterruptProcessingException', 'CancelledError')
diff --git a/freevideo_engine/comfy_host.py b/freevideo_engine/comfy_host.py
index 3a7f8f5..9f7d172 100644
--- a/freevideo_engine/comfy_host.py
+++ b/freevideo_engine/comfy_host.py
@@ -13,6 +13,7 @@ import time
from . import network, processes
from .comfy_bridge import installation
from .comfy_environment import isolated_environment
+from .environments import uv_file_arguments
from .locking import runtime_lock
from .monitoring import save
from .package_progress import PackageOutput
@@ -86,12 +87,13 @@ def _prepare(root, comfy, machine, run, download):
else:
log = directory / ('%02d-command.log' % count)
progress, child, success = LogProgress(log), None, False
+ command, cwd = uv_file_arguments(command, comfy if comfy.is_dir() else root)
ui.begin(key, label)
try:
with log.open('w', encoding='utf-8') as stream:
- with PackageOutput(list(map(str, command)), stream, environment) as output:
- child = processes.popen(list(map(str, command)), env=output.env,
- cwd=comfy if comfy.is_dir() else root, stdin=subprocess.DEVNULL,
+ with PackageOutput(command, stream, environment) as output:
+ child = processes.popen(command, env=output.env,
+ cwd=cwd, stdin=subprocess.DEVNULL,
stdout=output.stdout, stderr=subprocess.STDOUT, supervise=True, start_new_session=True)
output.spawned()
while child.poll() is None:
diff --git a/freevideo_engine/comfy_library.py b/freevideo_engine/comfy_library.py
index e073f78..595da37 100644
--- a/freevideo_engine/comfy_library.py
+++ b/freevideo_engine/comfy_library.py
@@ -2,7 +2,10 @@
import asyncio
import io
import json
+import os
import re
+import stat
+import time
from datetime import datetime, timezone
from pathlib import Path
@@ -123,6 +126,121 @@ def video_download(output_directory, identity):
return path, 'FreeVideo_%s_%s.mp4' % (stamp, identity.split('/')[-1][:12])
+# What a FreeVideo request writes into its own run folder. A creation whose
+# folder holds anything else, or any link, is not deleted at all.
+_OWNED_FILE = re.compile(r'video\.[A-Za-z0-9._-]+|prompt\.txt|workflow\.json|comfy-request\.json|generate\.log'
+ r'|media\.json|input-conditioning\.pt|(?:first|last)\.png|reference-\d{2}\.(?:png|mp4|wav)')
+_OWNED_FOLDER = 'video.artifacts'
+_REPARSE_POINT = 0x400
+
+
+class DeleteRefused(ValueError):
+ def __init__(self, reason, kept=()):
+ super().__init__(reason)
+ self.reason, self.kept = reason, list(kept)
+
+
+def _plain(path):
+ """A regular file or directory: never a symlink, junction or other reparse point."""
+ info = os.lstat(path)
+ if stat.S_ISLNK(info.st_mode) or getattr(info, 'st_file_attributes', 0) & _REPARSE_POINT:
+ return False
+ return stat.S_ISREG(info.st_mode) or stat.S_ISDIR(info.st_mode)
+
+
+def _foreign(folder):
+ """Names in a run folder that FreeVideo did not write, or that are links."""
+ kept = []
+ for entry in sorted(folder.iterdir()):
+ if not _plain(entry):
+ kept.append(entry.name)
+ elif entry.is_dir():
+ if entry.name != _OWNED_FOLDER:
+ kept.append(entry.name + '/')
+ continue
+ for parent, folders, files in os.walk(entry):
+ for name in folders + files:
+ if not _plain(Path(parent) / name):
+ kept.append((Path(parent) / name).relative_to(folder).as_posix())
+ elif not _OWNED_FILE.fullmatch(entry.name):
+ kept.append(entry.name)
+ return kept
+
+
+def _remove(folder):
+ """Remove a verified folder bottom-up; return bytes freed and paths left behind."""
+ freed, left = 0, []
+ for parent, folders, files in os.walk(folder, topdown=False):
+ for name in files:
+ path = Path(parent) / name
+ try:
+ if not _plain(path):
+ left.append(str(path)); continue
+ size = path.stat().st_size
+ path.unlink()
+ freed += size
+ except OSError:
+ left.append(str(path))
+ for name in folders:
+ try:
+ (Path(parent) / name).rmdir()
+ except OSError:
+ left.append(str(Path(parent) / name))
+ try:
+ Path(folder).rmdir()
+ except OSError:
+ left.append(str(folder))
+ return freed, left
+
+
+def delete_video(output_directory, identity):
+ """Delete one saved creation: its whole run folder, only when FreeVideo wrote everything in it.
+
+ The folder is first renamed into FreeVideo/.deleted, so it leaves the
+ library at once, or, when Windows still has a file open, nothing changes.
+ """
+ path = _video(output_directory, identity)
+ _report(path)
+ root = Path(output_directory).resolve() / 'FreeVideo'
+ folder = path.parent
+ if folder.parent.parent != root or not _plain(folder.parent) or not _plain(folder):
+ raise DeleteRefused('not-freevideo')
+ kept = _foreign(folder)
+ if kept:
+ raise DeleteRefused('foreign-files', kept)
+ # A second pass of this preview may be reading its first pass right now.
+ for state_path in root.glob('*/*/comfy-request.json'):
+ try:
+ state = json.loads(state_path.read_text(encoding='utf-8')) if state_path.stat().st_size <= 1024 * 1024 else {}
+ except (OSError, ValueError):
+ continue
+ source = state.get('upscaled_preview') if isinstance(state, dict) else None
+ if state.get('status') in ('starting', 'running') and isinstance(source, str) and Path(source).resolve() == path:
+ raise DeleteRefused('in-use')
+ trash = root / '.deleted'
+ trash.mkdir(exist_ok=True)
+ if not _plain(trash):
+ raise DeleteRefused('not-freevideo')
+ target = trash / ('%s-%d' % (identity.replace('/', '-'), time.time_ns()))
+ try:
+ folder.rename(target)
+ except OSError as error:
+ raise DeleteRefused('in-use') from error
+ freed, left = _remove(target)
+ # A result-cache row naming this creation can no longer be reused.
+ for index in (root / '.result-cache').glob('*.json'):
+ try:
+ if index.stat().st_size <= 1024 * 1024 and json.loads(index.read_text(encoding='utf-8')).get('id') == identity:
+ index.unlink()
+ except (OSError, ValueError, AttributeError):
+ pass
+ # Finish earlier deletions that Windows had kept open.
+ for earlier in trash.iterdir():
+ if earlier != target and _plain(earlier) and earlier.is_dir():
+ freed += _remove(earlier)[0]
+ return dict(id=identity, freed_bytes=freed, left=len(left))
+
+
def register():
from aiohttp import web
import folder_paths
@@ -176,6 +294,18 @@ def register():
return web.Response(body=data, content_type='image/jpeg',
headers={'Cache-Control': 'private, max-age=86400'})
+ @server.routes.post('/freevideo/library/delete')
+ async def delete(request):
+ try:
+ body = await request.json()
+ result = await asyncio.to_thread(delete_video, folder_paths.get_output_directory(),
+ body.get('id') if isinstance(body, dict) else None)
+ except DeleteRefused as error:
+ return web.json_response(dict(deleted=False, reason=error.reason, kept=error.kept[:20]), status=409)
+ except (OSError, ValueError, TypeError):
+ raise web.HTTPNotFound(text='Saved video unavailable') from None
+ return web.json_response(dict(result, deleted=True))
+
@server.routes.get('/freevideo/library/download')
async def download(request):
try:
diff --git a/freevideo_engine/comfy_nodes.py b/freevideo_engine/comfy_nodes.py
index 7645fe1..74dff8f 100644
--- a/freevideo_engine/comfy_nodes.py
+++ b/freevideo_engine/comfy_nodes.py
@@ -47,6 +47,12 @@ class FreeVideoGenerate(io.ComfyNode):
tooltip='Default: 3. Three steps use the independent refinement schedule. Changing sampling steps may reduce generation quality.'),
io.Boolean.Input('force_regenerate', display_name='Force regeneration', default=False, optional=True,
tooltip='Generate again even when an identical completed video is saved locally.'),
+ io.Boolean.Input('preview', display_name='Two-pass: preview first pass', default=False, optional=True,
+ tooltip='Part of two-pass acceleration: pause after the half-resolution first pass and save it as a preview. '
+ 'Run the second pass from the result when it looks right; it continues from the same latents.'),
+ io.String.Input('refine_from', display_name='Second pass of preview', default='', optional=True, advanced=True,
+ socketless=True, tooltip='Set by the Run second pass button: the preview video to continue. '
+ "The second pass uses the preview's own prompt, seed, size and inputs."),
],
outputs=[io.Video.Output('video'), io.String.Output('report', display_name='Report JSON')],
hidden=[io.Hidden.unique_id, io.Hidden.prompt, io.Hidden.extra_pnginfo],
@@ -57,8 +63,11 @@ class FreeVideoGenerate(io.ComfyNode):
def validate_inputs(cls, text, width, height, seconds, seed, **kwargs):
try:
comfy_bridge.validate_request(text, width, height, seconds, seed)
- from .two_pass import validate_steps
- validate_steps(kwargs.get('base_steps', 8), kwargs.get('refine_steps', 3), kwargs.get('two_pass', True))
+ if not kwargs.get('refine_from'):
+ from .two_pass import validate_steps
+ validate_steps(kwargs.get('base_steps', 8), kwargs.get('refine_steps', 3), kwargs.get('two_pass', True))
+ if kwargs.get('preview') and not kwargs.get('two_pass', True):
+ return 'Preview first needs two-pass acceleration.'
comfy_bridge.installation()
except (OSError, ValueError, KeyError) as error:
return str(error)
@@ -73,9 +82,10 @@ class FreeVideoGenerate(io.ComfyNode):
@classmethod
def execute(cls, text, width, height, seconds, seed, first=None, last=None,
references=None, loras=None, conditioning=None, media=None, two_pass=True,
- force_regenerate=False, base_steps=8, refine_steps=3):
+ force_regenerate=False, base_steps=8, refine_steps=3, preview=False, refine_from=''):
from .two_pass import validate_steps
- validate_steps(base_steps, refine_steps, two_pass)
+ if not refine_from:
+ validate_steps(base_steps, refine_steps, two_pass)
from .comfy_media import export
import folder_paths
import comfy.model_management as memory
@@ -87,7 +97,7 @@ class FreeVideoGenerate(io.ComfyNode):
prewarm = {}
result_reused = [False]
node_id = cls.hidden.unique_id
- bar = ProgressBar(base_steps + (refine_steps if two_pass else 0), node_id=node_id)
+ bar = ProgressBar(base_steps + (refine_steps if two_pass and not preview else 0), node_id=node_id)
last_message = [None]
last_count = [None]
def progress(message):
@@ -126,6 +136,10 @@ class FreeVideoGenerate(io.ComfyNode):
except (ImportError, AttributeError):
embed = True
graph = dict(prompt=getattr(cls.hidden, 'prompt', None), workflow=metadata.get('workflow')) if embed else None
+ if graph and refine_from:
+ # The upscaled video is the full request's result: embed that
+ # workflow, not one that points at this machine's preview file.
+ graph = comfy_bridge.without_upscale(graph, node_id, refine_from)
output_root = Path(folder_paths.get_output_directory()).resolve()
publish(PromptServer.instance, node_id, {'label': 'Preparing video', 'new_request': True})
try:
@@ -134,6 +148,7 @@ class FreeVideoGenerate(io.ComfyNode):
prompt_rewrite_report=rewrite_report,
encoder_prewarm=prewarm, force_regenerate=force_regenerate,
base_steps=base_steps, refine_steps=refine_steps, comfy_metadata=graph,
+ preview=bool(preview), refine_from=refine_from or None,
interrupted=memory.throw_exception_if_processing_interrupted, release_models=release,
export_inputs=lambda run, canvas: export(run, canvas, first=first, last=last,
references=references, loras=loras, conditioning=conditioning, assets=media))
@@ -153,9 +168,10 @@ class FreeVideoGenerate(io.ComfyNode):
publish(PromptServer.instance, node_id, {'label': 'Reused previous result' if result_reused[0] else 'Video saved', 'phase': 'complete',
'overall': {'status': 'complete', 'fraction': 1.}, 'result': preview['freevideo_summary'][0]})
measured = json.loads(report)
- encoder_path = measured.get('encoding', {}).get('encoder')
+ from .encoder_prewarm import warm_target
+ target = warm_target(measured)
from .resident_process import OWNER
- if encoder_path and not result_reused[0]:
+ if target and not result_reused[0]:
server = PromptServer.instance
client = server.client_id if server is not None else None
def busy():
@@ -164,13 +180,18 @@ class FreeVideoGenerate(io.ComfyNode):
if server is not None:
server.send_sync('freevideo_prewarm', dict(value, node=node_id), sid=client)
try:
- if OWNER.process is not None:
+ purpose, path = target
+ if purpose == 'upscale':
+ IDLE.schedule(path, output.with_suffix('.prewarm.json'),
+ ram_budget=measured.get('profile', {}).get('inference_ram_budget_gb', 0) * 1e9,
+ busy=busy, notify=warmed, purpose='upscale')
+ elif OWNER.process is not None:
root, machine = comfy_bridge.installation()
environment = comfy_bridge.engine_environment(root, comfy_bridge.source_root())
environment.update(FREEVIDEO_HOME=str(root), PYTHONPATH=str(comfy_bridge.source_root()))
OWNER.warm_encoder(output, machine['python'], environment, busy=busy, notify=warmed)
else:
- IDLE.schedule(encoder_path, output.with_suffix('.prewarm.json'),
+ IDLE.schedule(path, output.with_suffix('.prewarm.json'),
ram_budget=measured.get('profile', {}).get('inference_ram_budget_gb', 0) * 1e9,
busy=busy, notify=warmed)
except (OSError, RuntimeError):
diff --git a/freevideo_engine/decode_resume.py b/freevideo_engine/decode_resume.py
index 35ed891..76c4e6a 100644
--- a/freevideo_engine/decode_resume.py
+++ b/freevideo_engine/decode_resume.py
@@ -187,6 +187,79 @@ def save_refine_input(value, artifacts, torch):
temporary.unlink(missing_ok=True)
+def save_first_pass(value, artifacts, torch):
+ """Retain a preview's first pass atomically: the input of a later upscale."""
+ artifacts = Path(artifacts)
+ artifacts.mkdir(parents=True, exist_ok=True)
+ with tempfile.NamedTemporaryFile(prefix='first-pass.', suffix='.tmp', dir=artifacts, delete=False) as stream:
+ temporary = Path(stream.name)
+ try:
+ torch.save(value, temporary)
+ temporary.replace(artifacts / 'first-pass.pt')
+ finally:
+ temporary.unlink(missing_ok=True)
+
+
+def first_pass_complete(metrics):
+ """A finished preview retained its whole first pass, before the latent upscale."""
+ if (not isinstance(metrics, dict) or metrics.get('success') is not True or metrics.get('preview') is not True
+ or metrics.get('first_pass_checkpoint_complete') is not True
+ or not isinstance(metrics.get('refine_provenance'), dict)):
+ return False
+ plan, passes = metrics.get('sampling_plan'), metrics.get('sampling_passes')
+ if not isinstance(plan, dict) or not plan.get('enabled') or not isinstance(passes, list) or len(passes) != 1:
+ return False
+ first = passes[0]
+ if not isinstance(first, dict):
+ return False
+ steps, seconds = first.get('step_seconds'), first.get('sample_seconds')
+ return (isinstance(steps, list) and len(steps) == plan.get('base_steps')
+ and all(type(value) in (int, float) and math.isfinite(value) and value > 0 for value in steps)
+ and type(seconds) in (int, float) and math.isfinite(seconds) and seconds >= 0)
+
+
+def load_first_pass(request, torch):
+ """Load a finished preview's first pass on the CPU; the caller moves it to the GPU."""
+ resume = request['resume_first_pass']
+ if not isinstance(resume, dict) or any(not isinstance(resume.get(key), str) or not resume[key]
+ for key in ('input', 'metrics', 'request')):
+ raise ValueError('Invalid preview source')
+ paths = {key: Path(resume[key]).resolve() for key in ('input', 'metrics', 'request')}
+ if len({path.parent for path in paths.values()}) != 1 or len(set(paths.values())) != 3:
+ raise ValueError('Preview artifacts must belong to one finished preview')
+ prior = json.loads(paths['request'].read_text(encoding='utf-8'))
+ previous = json.loads(paths['metrics'].read_text(encoding='utf-8'))
+ if not isinstance(prior, dict) or not first_pass_complete(previous):
+ raise ValueError('Upscaling requires a finished preview')
+ provenance = refine_provenance(request)
+ if provenance != refine_provenance(prior) or provenance != previous['refine_provenance']:
+ raise ValueError('The preview does not match this request and its current inputs')
+ canvas, first = request['geometry'], request['sampling_plan']['first']
+ value = torch.load(str(paths['input']), map_location='cpu', weights_only=True)
+ if (not isinstance(value, dict) or type(value.get('seed')) is not int or value['seed'] != request['seed']
+ or value.get('geometry') != canvas or value.get('refine_provenance') != provenance):
+ raise ValueError('Preview first-pass provenance differs from its receipt')
+ width, height, frames, fps = first['width'], first['height'], first['frames'], canvas['fps']
+ if (any(type(n) is not int or n <= 0 for n in (width, height, frames, fps))
+ or width % 32 or height % 32 or frames % 17 != 5 or fps != 24 or frames != canvas['frames']):
+ raise ValueError('Invalid preview first-pass geometry')
+ # The first pass samples the smaller planned canvas: video C=24 with
+ # stride 16; stereo audio C=32 at 40/s for the same duration.
+ shapes = {'video': (1, 24, (frames - 5) // 17 * 5 + 2, height // 16, width // 16),
+ 'audio': (2, 32, round(frames / fps * 40))}
+ for key, shape in shapes.items():
+ tensor = value.get(key)
+ if (not isinstance(tensor, torch.Tensor) or tensor.device.type != 'cpu'
+ or tuple(tensor.shape) != shape or not tensor.is_floating_point()
+ or not bool(torch.isfinite(tensor).all())):
+ raise ValueError('Preview first-pass %s latents are malformed or non-finite' % key)
+ receipt = dict(first_pass_reused=True, first_pass_source='preview', refine_provenance=provenance,
+ first_pass_source_output=prior.get('output'),
+ first_pass_scope='Reused the first pass of a finished preview; its times and peaks '
+ 'belong to that preview.')
+ return value['video'], value['audio'], copy.deepcopy(previous['sampling_passes'][0]), receipt
+
+
def refine_checkpoint_complete(metrics):
"""A failed attempt retained its full first pass, latent upscale and crop."""
if (not isinstance(metrics, dict) or metrics.get('success') is not False
diff --git a/freevideo_engine/diagnostic_resources.py b/freevideo_engine/diagnostic_resources.py
index 0b9515c..95ed627 100644
--- a/freevideo_engine/diagnostic_resources.py
+++ b/freevideo_engine/diagnostic_resources.py
@@ -119,6 +119,8 @@ def encoder_prewarm(value):
for key in ('gpu_preload', 'worker_released', 'receipt_available'):
if type(value.get(key)) is bool:
result[key] = value[key]
+ if value.get('purpose') in ('encoder', 'upscale'):
+ result['purpose'] = value['purpose']
return result
diff --git a/freevideo_engine/encoder_prewarm.py b/freevideo_engine/encoder_prewarm.py
index 775318f..7fa39ba 100644
--- a/freevideo_engine/encoder_prewarm.py
+++ b/freevideo_engine/encoder_prewarm.py
@@ -1,8 +1,9 @@
-"""Idle, interruptible encoder file-cache prefetch for interactive sessions.
+"""Idle, interruptible model file-cache prefetch for interactive sessions.
-The native encoder maps its checkpoint. Reading it while idle warms the same OS
-pages without keeping another Torch process, private model copy or CUDA context.
-The OS can reclaim all warmed pages. This is not GPU model residency.
+The native encoder maps its checkpoint, and the video engine maps the
+transformer blocks it does not keep on the GPU. Reading them while idle warms
+the same OS pages without keeping another Torch process, private model copy or
+CUDA context. The OS can reclaim all warmed pages. This is not GPU residency.
"""
import os
from pathlib import Path
@@ -29,6 +30,21 @@ def plan(size, available, *, ram_budget=None, complete=True):
else 'Leave memory for active applications; encoder loads on demand'))
+def warm_target(report):
+ """What an idle session warms after this saved video: (purpose, files) or None.
+
+ After a preview the likely next request is its upscale, which encodes
+ nothing, so warm the files it reads; the encoder's pages would push those
+ out of a small OS file cache. After any other video, warm the encoder.
+ """
+ preview = report.get('preview') or {}
+ if preview:
+ files = (preview.get('upscale_reads') or {}).get('files')
+ return ('upscale', files) if files else None
+ encoder = (report.get('encoding') or {}).get('encoder')
+ return ('encoder', encoder) if encoder else None
+
+
def available_memory():
memory = system_memory()
group = cgroup_memory()
@@ -41,38 +57,44 @@ def available_memory():
def prefetch(path, selected, cancelled, *, busy=None, notify=None, availability=available_memory):
- path = Path(path)
+ """Read one checkpoint, or several files in the order given, into the OS file cache."""
+ paths = [Path(item) for item in path] if isinstance(path, (list, tuple)) else [Path(path)]
result = dict(selected, state='skipped', done_bytes=0)
started = time.monotonic()
if not selected['enabled']:
return result
- initial = path.stat()
result['state'] = 'reading'
buffer = bytearray(CHUNK)
last = 0.
- with path.open('rb', buffering=0) as stream:
- while result['done_bytes'] < selected['bytes']:
- if cancelled.is_set() or busy and busy():
- result.update(state='cancelled', reason='A new request takes priority')
- break
- available, complete = availability()
- if not complete or available < selected['system_reserve_bytes'] + CHUNK:
- result.update(state='released', reason='Live RAM pressure; warmed pages remain reclaimable')
- break
- count = stream.readinto(buffer)
- if not count:
- raise ValueError('Encoder checkpoint changed during prefetch')
- result['done_bytes'] += count
- elapsed = time.monotonic() - started
- result.update(elapsed_seconds=elapsed, bytes_per_second=result['done_bytes'] / max(.001, elapsed))
- if notify and elapsed - last >= 1:
- notify(dict(result))
- last = elapsed
- else:
- current = path.stat()
- if (current.st_size, current.st_mtime_ns) != (initial.st_size, initial.st_mtime_ns):
- raise ValueError('Encoder checkpoint changed during prefetch')
- result.update(state='ready', reason='Checkpoint pages read; native loader can reuse them if still cached')
+ for item in paths:
+ initial = item.stat()
+ remaining = initial.st_size
+ with item.open('rb', buffering=0) as stream:
+ while remaining > 0:
+ if cancelled.is_set() or busy and busy():
+ result.update(state='cancelled', reason='A new request takes priority')
+ break
+ available, complete = availability()
+ if not complete or available < selected['system_reserve_bytes'] + CHUNK:
+ result.update(state='released', reason='Live RAM pressure; warmed pages remain reclaimable')
+ break
+ count = stream.readinto(buffer)
+ if not count:
+ raise ValueError('A model file changed during prefetch')
+ remaining -= count
+ result['done_bytes'] += count
+ elapsed = time.monotonic() - started
+ result.update(elapsed_seconds=elapsed, bytes_per_second=result['done_bytes'] / max(.001, elapsed))
+ if notify and elapsed - last >= 1:
+ notify(dict(result))
+ last = elapsed
+ if result['state'] != 'reading':
+ break
+ current = item.stat()
+ if (current.st_size, current.st_mtime_ns) != (initial.st_size, initial.st_mtime_ns):
+ raise ValueError('A model file changed during prefetch')
+ else:
+ result.update(state='ready', reason='File pages read; the next load reuses them while they stay cached')
result['elapsed_seconds'] = time.monotonic() - started
return result
@@ -92,7 +114,9 @@ class IdlePrewarmer:
from .encoder_diagnostics import prewarm_receipt
return prewarm_receipt(self.report, gpu=False)
- def schedule(self, path, report, *, ram_budget=None, busy=None, notify=None):
+ def schedule(self, path, report, *, ram_budget=None, busy=None, notify=None, purpose='encoder'):
+ """Warm the encoder checkpoint, or the files an upscale of a preview reads."""
+ paths = [Path(item) for item in path] if isinstance(path, (list, tuple)) else [Path(path)]
with self.guard:
self.stop()
if self.thread and self.thread.is_alive():
@@ -101,26 +125,31 @@ class IdlePrewarmer:
self.report = Path(report)
cancel = self.cancelled
def work():
- result = {'state': 'cancelled', 'reason': 'A new request takes priority'}
+ result = {'state': 'cancelled', 'reason': 'A new request takes priority', 'purpose': purpose}
try:
# Return the completed video to the user before doing work.
if cancel.wait(2.) or busy and busy():
return
available, complete = available_memory()
- selected = plan(Path(path).stat().st_size, available, ram_budget=ram_budget, complete=complete)
+ selected = plan(sum(item.stat().st_size for item in paths), available,
+ ram_budget=ram_budget, complete=complete)
+ selected['purpose'] = purpose
+ if not selected['enabled'] and purpose == 'upscale':
+ selected['reason'] = 'Leave memory for active applications; the upscale reads its files on demand'
+ # One switch for all idle file warming.
if os.environ.get('FREEVIDEO_ENCODER_PREWARM', 'auto').lower() in ('0', 'off', 'false'):
selected.update(enabled=False, reason='Disabled by FREEVIDEO_ENCODER_PREWARM')
- result = prefetch(path, selected, cancel, busy=busy, notify=notify)
+ result = prefetch(paths, selected, cancel, busy=busy, notify=notify)
except Exception as error:
- result = {'state': 'failed', 'reason': str(error)}
+ result = {'state': 'failed', 'reason': str(error), 'purpose': purpose}
finally:
try:
- save(report, dict(result, checkpoint=str(path), completed_epoch=time.time()))
+ save(report, dict(result, files=[str(item) for item in paths], completed_epoch=time.time()))
if notify:
notify(result)
except (OSError, RuntimeError):
pass # Optional idle preparation cannot fail a saved video.
- self.thread = threading.Thread(target=work, name='FreeVideo encoder prefetch', daemon=True)
+ self.thread = threading.Thread(target=work, name='FreeVideo %s prefetch' % purpose, daemon=True)
self.thread.start()
return True
diff --git a/freevideo_engine/environments.py b/freevideo_engine/environments.py
index 35f9560..6fdb6a3 100644
--- a/freevideo_engine/environments.py
+++ b/freevideo_engine/environments.py
@@ -1,4 +1,5 @@
"""Pinned installation layouts; model lifetimes are independent of venv count."""
+import os
from pathlib import Path
import platform
@@ -19,6 +20,47 @@ def constraints_file(name, system=None):
return Path(__file__).resolve().parent.parent / 'constraints' / (prefix + name + '.txt')
+# uv reads each value of these options as a space-separated list of files, even
+# when it arrives as one quoted argument: `-c "D:\Comfy UI\pins.txt"` names
+# `D:\Comfy` and `UI\pins.txt` and stops with "File not found" (exit code 2).
+# Requirement files, editables, wheels and --python keep their spaces.
+UV_FILE_LIST_OPTIONS = frozenset(('-c', '--constraint', '--constraints', '-b', '--build-constraint',
+ '--build-constraints', '--override', '--overrides'))
+
+
+def uv_file_arguments(command, cwd=None):
+ """Return (command, cwd) for which uv reads constraint files in folders with spaces.
+
+ Such a file is named relative to its folder, which becomes uv's working
+ directory. FreeVideo passes every other path to uv as an absolute path.
+ """
+ command = [str(item) for item in command]
+ if not command or Path(command[0]).name.lower() not in ('uv', 'uv.exe') or command[1:2] != ['pip']:
+ return command, cwd
+ values = {}
+ for index, item in enumerate(command):
+ option, equals, value = item.partition('=')
+ if item in UV_FILE_LIST_OPTIONS and index + 1 < len(command):
+ values[index + 1] = ('', command[index + 1])
+ elif equals and option.startswith('--') and option in UV_FILE_LIST_OPTIONS:
+ values[index] = (option + '=', value)
+ spaced = {index: row for index, row in values.items() if ' ' in row[1]}
+ if not spaced:
+ return command, cwd
+ base = Path(cwd or os.getcwd())
+ folder = (base / next(iter(spaced.values()))[1]).parent
+ result = list(command)
+ for index, (prefix, value) in spaced.items():
+ try:
+ relative = os.path.relpath(base / value, folder)
+ except ValueError: # Another Windows drive; uv reports the original path.
+ return command, cwd
+ if ' ' in relative:
+ return command, cwd
+ result[index] = prefix + relative
+ return result, str(folder)
+
+
def bootstrap_versions(value, system=None):
result = dict(value, target_system=system or platform.system())
if result['target_system'] == 'Darwin':
diff --git a/freevideo_engine/generate.py b/freevideo_engine/generate.py
index 0b9a5e6..f1e486d 100644
--- a/freevideo_engine/generate.py
+++ b/freevideo_engine/generate.py
@@ -302,6 +302,10 @@ def _run(args):
requested_steps = getattr(args, 'base_steps', None)
refine_steps = getattr(args, 'refine_steps', 3)
two_pass = getattr(args, 'two_pass', True)
+ preview = getattr(args, 'preview', False) is True
+ from .preview import validate_flags, source as preview_source
+ validate_flags(preview, getattr(args, 'refine_from', None), two_pass)
+ refine_source = preview_source(args.refine_from) if getattr(args, 'refine_from', None) else None
if requested_steps is not None:
validate_steps(requested_steps, refine_steps, two_pass)
if requested_steps != 8:
@@ -446,9 +450,15 @@ def _run(args):
PYTORCH_ALLOC_CONF=allocator, PYTORCH_CUDA_ALLOC_CONF=allocator)
env[LOCK_ENV] = str(descriptor)
condition = temporary / 'conditioning.pt'
- if args.conditioning:
+ if refine_source:
+ # The preview's own conditioning: the exact file its first pass used.
+ condition = Path(refine_source['conditioning'])
+ if refine_source['report'].get('encoding'):
+ report['encoding'] = dict(refine_source['report']['encoding'], reused_from_preview=True)
+ print(json.dumps({'event': 'conditioning_reused', 'source': 'preview'}), flush=True)
+ elif args.conditioning:
shutil.copyfile(args.conditioning, condition)
- if args.prompt_file:
+ if args.prompt_file and not refine_source:
encoding = {'prompt': args.prompt_file.read_text(encoding='utf-8'), 'encoder': args.encoder,
'comfy_root': str(args.comfy_root or comfy_root()),
'model_paths': str(args.model_paths.resolve()) if args.model_paths else None,
@@ -540,9 +550,13 @@ def _run(args):
sampling_plan = plan_sampling(canvas, two_pass, profile['engine'].get('task', 't2va'),
base_steps=base_steps, refine_steps=refine_steps)
report['sampling_plan'] = sampling_plan
+ if preview:
+ report['preview'] = dict(geometry=dict(sampling_plan['first'], fps=canvas['fps']))
+ if refine_source:
+ report['upscaled_preview'] = dict(output=refine_source['output'])
if sampling_plan['enabled'] or sampling_plan['version'] == 2:
canvas['sampling_plan'] = sampling_plan
- print(json.dumps(dict(event='sampling_plan', **sampling_plan)), flush=True)
+ print(json.dumps(dict(event='sampling_plan', **sampling_plan, **({'preview': True} if preview else {}))), flush=True)
save(destination.with_suffix('.request.json'), report)
upscaler_checkpoint = None
if sampling_plan.get('upscaler_sha256'):
@@ -624,18 +638,25 @@ def _run(args):
'sampling_plan': sampling_plan, 'upscaler_checkpoint': upscaler_checkpoint,
'automatic_pass_cache': not bool(args.profile),
'metrics': str(destination.with_suffix('.engine.json'))}
+ if preview:
+ request['preview'] = True
if allocator_limit_bytes is not None:
request['allocator_limit_bytes'] = allocator_limit_bytes
report['benchmark_limit'] = {'pytorch_allocator_bytes': request['allocator_limit_bytes'],
'scope': 'PyTorch CUDA allocator only; other CUDA allocations and whole-device usage are measured separately.'}
request_path = temporary / 'video.json'
resume_decode = resume_refine = None
+ # An upscale reuses its preview's first pass unless a later attempt
+ # retains a newer checkpoint, or the current inputs no longer match.
+ preview_pass = ({key: refine_source[key] for key in ('input', 'metrics', 'request')}
+ if refine_source else None)
# Keep explicit dependency overlays while keeping the snapshotted engine first.
engine_env = dict(env, PYTHONPATH=os.pathsep.join((str(repo), str(vdn_root()),
env.get('PYTHONPATH', ''))))
def persist():
save(destination.with_suffix('.request.json'), report)
def launch(selected, attempt):
+ nonlocal preview_pass
selected_env = inference_environment(engine_env, selected['inference_ram_budget_gb'] * 1e9)
request.update(engine_options=selected['engine'], decoder_options=selected['decoder'],
ram_budget_bytes=round(selected['inference_ram_budget_gb'] * 1e9),
@@ -667,6 +688,21 @@ def _run(args):
request['resume_refine'] = dict(resume_refine)
else:
request.pop('resume_refine', None)
+ request.pop('resume_first_pass', None)
+ if resume_decode is None and resume_refine is None and preview_pass is not None:
+ from .decode_resume import refine_provenance
+ receipt = json.loads(Path(preview_pass['metrics']).read_text(encoding='utf-8'))
+ if refine_provenance(request) == receipt.get('refine_provenance'):
+ request['resume_first_pass'] = dict(preview_pass)
+ report['upscaled_preview']['first_pass_reused'] = True
+ else:
+ # Other models, settings or math: the preview's first pass
+ # is not this request's. Sample it again instead.
+ preview_pass = None
+ report['upscaled_preview'].update(first_pass_reused=False,
+ reason='The preview no longer matches the current models or settings; '
+ 'its first pass was sampled again.')
+ print(json.dumps(dict(event='preview_mismatch')), flush=True)
save(request_path, request)
print(json.dumps({'event': 'decode_resume' if resume_decode else 'video_start', 'resource_attempt': attempt,
'engine': selected['engine'], 'decoder': selected['decoder']}), flush=True)
@@ -748,7 +784,8 @@ def _run(args):
retry_state = dict(attempt=index + 2, max_attempts=getattr(args, 'resource_retries', 2) + 1,
failed_phase=row.get('phase'), kind=row['failure']['kind'],
reuse_sampling=resume_decode is not None,
- reuse_first_pass=resume_decode is None and resume_refine is not None)
+ reuse_first_pass=resume_decode is None and (resume_refine is not None
+ or preview_pass is not None))
row['recovery'] = dict(decision, next_profile=selected, **retry_state)
persist()
print(json.dumps(dict(event='resource_retry', retained=row['retained'], retry=retry_state)), flush=True)
@@ -761,6 +798,11 @@ def _run(args):
archive=archive, report=report, persist=persist, on_retry=retry,
automatic=not args.profile or getattr(args, 'recover_placement', False),
max_retries=getattr(args, 'resource_retries', 2))
+ if preview:
+ # An upscale validates this receipt beside the retained first pass.
+ shutil.copyfile(destination.with_suffix('.engine.json'), artifacts / 'first-pass.engine.json')
+ from .preview import upscale_reads
+ report['preview']['upscale_reads'] = upscale_reads(request['cache'], report['video'], upscaler_checkpoint)
report['success'] = True
except BaseException as error:
from .diagnostic_resources import exception_details
diff --git a/freevideo_engine/macos_generate.py b/freevideo_engine/macos_generate.py
index d0781f1..f5946b9 100644
--- a/freevideo_engine/macos_generate.py
+++ b/freevideo_engine/macos_generate.py
@@ -10,6 +10,7 @@ import math
import os
from pathlib import Path
import platform
+import shutil
import sys
import time
import traceback
@@ -106,6 +107,10 @@ def run(args):
from .support_report import write as write_debug, code_identity
from .two_pass import plan
require_native()
+ from .preview import validate_flags, source as preview_source
+ preview = getattr(args, 'preview', False) is True
+ validate_flags(preview, getattr(args, 'refine_from', None), args.two_pass)
+ refine_source = preview_source(args.refine_from, native=True) if getattr(args, 'refine_from', None) else None
output = Path(args.out).expanduser().resolve()
artifacts = output.with_suffix('.artifacts')
if any(p.exists() for p in (output, artifacts, output.with_suffix('.request.json'))):
@@ -148,6 +153,10 @@ def run(args):
report.update(geometry=canvas, sampling_plan=sampling, profile=dict(policy=resource,
engine=dict(device_backend='mps', task=task, attention='mps', steps=sampling['base_steps'])),
encoder_mode='native H3 encoder library child' if args.prompt_file else 'preencoded shared conditioning')
+ if preview:
+ report['preview'] = dict(geometry=dict(sampling['first'], fps=canvas['fps']))
+ if refine_source:
+ report['upscaled_preview'] = dict(output=refine_source['output'])
env = environment(data_root(), dict(os.environ, PYTHONUNBUFFERED='1', OMP_NUM_THREADS='2', MKL_NUM_THREADS='2'))
with runtime_lock() as descriptor, awake() as idle_sleep_prevented:
report['idle_sleep_prevented'] = idle_sleep_prevented
@@ -171,7 +180,20 @@ def run(args):
prepared_cache = Path(prepared['cache']).resolve()
report['lora'] = prepared['report']
conditioning = Path(args.conditioning).resolve() if args.conditioning else artifacts / 'conditioning.pt'
- if args.prompt_file:
+ if preview and args.conditioning:
+ # Keep the preview self-contained for its later upscale.
+ shutil.copyfile(conditioning, artifacts / 'conditioning.pt')
+ conditioning = artifacts / 'conditioning.pt'
+ if refine_source:
+ # The preview's own conditioning: the exact file its first pass used.
+ conditioning = Path(refine_source['conditioning'])
+ prior = refine_source['report']
+ if prior.get('encoding'):
+ report['encoding'] = dict(prior['encoding'], reused_from_preview=True)
+ task = prior.get('profile', {}).get('engine', {}).get('task', task)
+ report['profile']['engine']['task'] = task
+ print(json.dumps(dict(event='conditioning_reused', source='preview')), flush=True)
+ if args.prompt_file and not refine_source:
report['phase'] = 'encoding'
save(output.with_suffix('.request.json'), report)
prompt = Path(args.prompt_file).read_text(encoding='utf-8')
@@ -211,17 +233,24 @@ def run(args):
raise FileNotFoundError('Conditioning is missing: ' + str(conditioning))
report['phase'] = 'engine'
save(output.with_suffix('.request.json'), report)
- print(json.dumps(dict(event='sampling_plan', **sampling)), flush=True)
- report['resources'] = execute(dict(phase='engine', output=str(output), artifacts=str(artifacts),
+ print(json.dumps(dict(event='sampling_plan', **sampling, **({'preview': True} if preview else {}))), flush=True)
+ engine = dict(phase='engine', output=str(output), artifacts=str(artifacts),
conditioning=str(conditioning), cache=str(prepared_cache),
base=base, task=task,
checkpoint=str(Path(args.checkpoint or checkpoint_path()).resolve()),
- seed=args.seed, canvas=canvas, sampling=sampling, resources=resource), 'engine')
+ seed=args.seed, canvas=canvas, sampling=sampling, resources=resource)
+ if preview:
+ engine['preview'] = True
+ if refine_source:
+ engine['refine_from'] = {key: refine_source[key] for key in ('input', 'metrics', 'request', 'output')}
+ report['resources'] = execute(engine, 'engine')
metrics = json.loads(output.with_suffix('.engine.json').read_text())
if (metrics.get('success') is not True or metrics.get('sampling_plan') != sampling
- or len(metrics.get('step_seconds', [])) != sampling['total_steps']
+ or len(metrics.get('step_seconds', [])) != (sampling['base_steps'] if preview else sampling['total_steps'])
or not output.is_file() or output.stat().st_size == 0):
raise RuntimeError('Native worker did not complete the requested video')
+ if refine_source:
+ report['upscaled_preview'].update(metrics.get('upscaled_preview') or {})
report.update(success=True, phase='complete', video=metrics)
return output
except BaseException as error:
diff --git a/freevideo_engine/macos_stages.py b/freevideo_engine/macos_stages.py
index 664c829..73b23b0 100644
--- a/freevideo_engine/macos_stages.py
+++ b/freevideo_engine/macos_stages.py
@@ -76,12 +76,63 @@ def retain_joined_progress(result, observed):
result['phase'] = observed.get('sampling_phase', result['phase'])
+def first_pass_identity(value):
+ """What a native first pass depends on: inputs, models, seed and sampling plan.
+
+ Compute partitions change floating-point reduction order, not the math, so
+ they are excluded, as for CUDA's retained first passes.
+ """
+ from .decode_resume import _digest
+ cache = Path(value['cache'])
+ return dict(seed=value['seed'], canvas=value['canvas'], sampling=value['sampling'],
+ task=value.get('task', 't2va'), base=str(Path(value['base']).resolve()),
+ checkpoint=str(Path(value['checkpoint']).resolve()), cache=str(cache.resolve()),
+ cache_manifest_sha256=_digest(cache / 'manifest.json'),
+ conditioning=str(Path(value['conditioning']).resolve()),
+ conditioning_sha256=_digest(value['conditioning']))
+
+
+def reuse_preview(value, artifacts, result):
+ """Start an upscale from its preview's first pass when it still matches."""
+ import shutil
+ source = value['refine_from']
+ receipt = json.loads(Path(source['metrics']).read_text(encoding='utf-8'))
+ passes = receipt.get('sampling_passes')
+ if (receipt.get('preview') is not True or receipt.get('first_pass_identity') != first_pass_identity(value)
+ or not isinstance(passes, list) or len(passes) != 1
+ or len(passes[0].get('step_seconds', [])) != value['sampling']['base_steps']):
+ result['upscaled_preview'] = dict(first_pass_reused=False,
+ reason='The preview no longer matches the current models or settings; its first pass was sampled again.')
+ print(json.dumps(dict(event='preview_mismatch')), flush=True)
+ return False
+ target = artifacts / 'first-pass.pt'
+ try:
+ os.link(source['input'], target)
+ except OSError:
+ shutil.copyfile(source['input'], target)
+ # The same first-pass receipt as a request that sampled it, naming its source.
+ save(artifacts / 'first-pass.json', dict(sampling_plan=value['sampling'], sampling_passes=list(passes),
+ first_pass_source='preview', preview_output=source.get('output')))
+ result.update(sampling_passes=list(passes), first_pass_reused=True, first_pass_source='preview',
+ upscaled_preview=dict(first_pass_reused=True, output=source.get('output')))
+ print(json.dumps(dict(event='first_pass_reused', source='preview', completed_steps=value['sampling']['base_steps'],
+ total=value['sampling']['total_steps'])), flush=True)
+ return True
+
+
def run(value):
output, artifacts = Path(value['output']), Path(value['artifacts'])
sampling = value['sampling']
path = output.with_suffix('.engine.json')
- result = dict(success=False, device_backend='mps', phase='load', geometry=value['canvas'],
+ preview = value.get('preview') is True
+ geometry = value['canvas']
+ if preview:
+ # A preview stops after the first pass and decodes it at that canvas.
+ geometry = dict(geometry, width=sampling['first']['width'], height=sampling['first']['height'])
+ result = dict(success=False, device_backend='mps', phase='load', geometry=geometry,
sampling_plan=sampling, stages={}, sampling_passes=[])
+ if preview:
+ result['preview'] = True
started = time.monotonic()
def interrupted(signum, frame):
raise KeyboardInterrupt('Native stage interrupted by signal ' + str(signum))
@@ -90,6 +141,10 @@ def run(value):
lifetime = sampling_lifetime(value['resources'])
with runtime_lock() as descriptor:
stages = ['first-pass', 'upscale', 'refinement', 'decode'] if sampling['enabled'] else ['first-pass', 'decode']
+ if preview:
+ stages = ['first-pass', 'decode']
+ elif value.get('refine_from') and reuse_preview(value, artifacts, result):
+ stages = ['upscale', 'refinement', 'decode']
joined = False
for stage in stages:
if joined and stage in ('upscale', 'refinement'):
@@ -136,8 +191,11 @@ def run(value):
elif stage in ('first-pass', 'refinement'):
result['sampling_passes'].append(observed['sampling'])
if stage == 'first-pass' and sampling['enabled']:
- save(artifacts / 'first-pass.json', dict(sampling_plan=sampling,
- sampling_passes=result['sampling_passes']))
+ receipt = dict(sampling_plan=sampling, sampling_passes=result['sampling_passes'])
+ if preview:
+ # An upscale of this preview checks it still matches.
+ receipt.update(preview=True, first_pass_identity=first_pass_identity(value))
+ save(artifacts / 'first-pass.json', receipt)
elif stage == 'upscale':
result['latent_upscale'] = observed['upscale']
else:
@@ -151,7 +209,7 @@ def run(value):
sample_seconds=sum(s['work_seconds'] for k, s in completed.items() if k != 'decode'),
latent_save_seconds=sum(s.get('latent_save_seconds', 0.) for s in completed.values()),
step_seconds=[t for row in result['sampling_passes'] for t in row['step_seconds']])
- if (len(result['step_seconds']) != sampling['total_steps'] or
+ if (len(result['step_seconds']) != (sampling['base_steps'] if preview else sampling['total_steps']) or
not output.is_file() or output.stat().st_size == 0):
raise RuntimeError('Native stages did not produce the complete requested video')
result.update(success=True, phase='complete')
@@ -235,7 +293,9 @@ def worker(value):
task=value.get('task', 't2va'), canvas=value['canvas']) as engine:
result['load_seconds'] = engine.load_seconds
tick = time.monotonic()
- admission = joined_admission(engine, value) if not refining else dict(selected=False)
+ # A preview stops after the first pass; joining would also refine.
+ admission = (joined_admission(engine, value) if not refining and value.get('preview') is not True
+ else dict(selected=False))
result['sampling_admission'] = admission
if admission['selected']:
result.update(joined_sampling=True, sampling_passes=[], latent_save_seconds=0.)
@@ -263,7 +323,8 @@ def worker(value):
else:
video, audio, measured = engine.sample(value['conditioning'],
value['seed'] + (sampling['restart_seed_offset'] if refining else 0),
- **sampling['second' if refining else 'first'], progress_total=sampling['total_steps'],
+ **sampling['second' if refining else 'first'],
+ progress_total=sampling['base_steps'] if value.get('preview') is True else sampling['total_steps'],
progress_offset=sampling['base_steps'] if refining else 0,
**(dict(initial_latents=(initial['video'], initial['audio']), refine_steps=sampling['refine_steps'],
refine_schedule=sampling.get('refine_schedule'))
@@ -290,7 +351,9 @@ def worker(value):
name = 'upscaled.pt'
elif phase == 'decode':
from .macos_decode import decode_to_file
- initial = torch.load(artifacts / 'latents.pt', map_location='cpu', weights_only=True)
+ # A preview decodes its first pass; the upscale input stays for later.
+ source = 'first-pass.pt' if value.get('preview') is True else 'latents.pt'
+ initial = torch.load(artifacts / source, map_location='cpu', weights_only=True)
tick = time.monotonic()
result['decode'] = decode_to_file(initial['video'], initial['audio'], value['output'],
base=value['base'], artifacts_dir=artifacts,
diff --git a/freevideo_engine/preview.py b/freevideo_engine/preview.py
new file mode 100644
index 0000000..c1908a2
--- /dev/null
+++ b/freevideo_engine/preview.py
@@ -0,0 +1,67 @@
+"""Previews: a two-pass request that stops after its half-resolution first pass.
+
+The first pass is kept beside the preview MP4, so an upscale continues from the
+same latents: the learned latent upscale, the refinement steps and a full
+decode, as the uninterrupted request would have run them. Torch-free.
+"""
+import json
+from pathlib import Path
+
+# CUDA keeps the worker request and a copy of its receipt next to the latents;
+# the decode_resume loaders require all three in one folder.
+CUDA_FILES = ('first-pass.pt', 'first-pass.engine.json', 'video.json', 'conditioning.pt')
+# Native Mac stages keep their own first-pass receipt and engine request.
+MAC_FILES = ('first-pass.pt', 'first-pass.json', 'engine-request.json', 'conditioning.pt')
+
+
+def validate_flags(preview, refine_from, two_pass):
+ if preview and refine_from:
+ raise ValueError('Choose either a preview or an upscale of one, not both')
+ if (preview or refine_from) and not two_pass:
+ raise ValueError('Previews and their upscale use two-pass generation')
+
+
+def source(output, native=False):
+ """The retained first pass of a finished preview MP4, or a clear error."""
+ output = Path(output).expanduser().resolve()
+ if output.suffix.lower() != '.mp4' or not output.is_file():
+ raise ValueError('Choose the preview MP4 to upscale')
+ folder = output.with_suffix('.artifacts')
+ names = MAC_FILES if native else CUDA_FILES
+ missing = [name for name in names if not (folder / name).is_file()]
+ if missing:
+ raise ValueError('This preview no longer has its first pass (%s). Generate the preview again.'
+ % ', '.join(missing))
+ report = json.loads(output.with_suffix('.request.json').read_text(encoding='utf-8'))
+ if not isinstance(report, dict) or report.get('success') is not True or not report.get('preview'):
+ raise ValueError('This video is not a finished preview')
+ if not native:
+ from .decode_resume import first_pass_complete
+ if not first_pass_complete(json.loads((folder / 'first-pass.engine.json').read_text(encoding='utf-8'))):
+ raise ValueError('This preview did not keep a complete first pass')
+ result = dict(output=str(output), report=report, conditioning=str(folder / 'conditioning.pt'),
+ input=str(folder / 'first-pass.pt'))
+ if native:
+ result.update(metrics=str(folder / 'first-pass.json'), request=str(folder / 'engine-request.json'))
+ else:
+ result.update(metrics=str(folder / 'first-pass.engine.json'), request=str(folder / 'video.json'))
+ return result
+
+
+def upscale_reads(cache, video, upscaler_checkpoint):
+ """Files the upscale of this preview reads from disk, in the order it needs them.
+
+ Transformer blocks the engine keeps on the GPU are not read again; the
+ others are mapped from the prepared cache at every step, and a decode in
+ between can push them out of a small OS file cache. The latent upscaler
+ is read once. An idle session can warm exactly these files.
+ """
+ resident = ((video or {}).get('config') or {}).get('resident_blocks')
+ files = []
+ if type(resident) is int and resident >= 0:
+ blocks = sorted(Path(cache, 'blocks').glob('[0-9][0-9].safetensors'), key=lambda path: int(path.stem))
+ files += [path for path in blocks if int(path.stem) >= resident]
+ if upscaler_checkpoint:
+ files.append(Path(upscaler_checkpoint))
+ files = [path for path in files if path.is_file()]
+ return dict(files=[str(path) for path in files], bytes=sum(path.stat().st_size for path in files))
diff --git a/freevideo_engine/release_notes.json b/freevideo_engine/release_notes.json
index d5b713a..ac2c012 100644
--- a/freevideo_engine/release_notes.json
+++ b/freevideo_engine/release_notes.json
@@ -1,18 +1,28 @@
{
"schema": 1,
- "product_version": "0.3.1",
+ "product_version": "0.3.2",
"en": {
- "summary": "Fixes the int8 upgrade on older installations and importing the Environment ZIP.",
+ "summary": "Adds Preview first to two-pass sampling, deleting in Your creations and prompt box improvements (the example prompt as a placeholder, the last prompt restored on reopening), and fixes installing into folders with spaces.",
"changes": [
- "Older installations could not upgrade to the int8 model: the upgrade stopped with \"The download is larger than shown\". The size shown now includes every file the upgrade downloads, such as the sampling tables, so the upgrade goes through.",
- "Importing the Environment ZIP failed with \"No such file or directory: windows-portable.txt\". It imports again."
+ "Two-pass sampling can stop at a quick half-resolution preview: choose Two-pass: Preview first. When a preview looks right, choose Run second pass; the finished video is identical to generating it directly. Make several previews in a batch and pick one in Your creations.",
+ "Your creations can delete videos you no longer need, after a confirmation.",
+ "The example prompt is shown greyed out as a placeholder. Typing replaces it without clearing the box first, and a box that was never typed in generates the example. A box emptied after typing stays empty.",
+ "Prompt enhancement rewrites the prompt that was typed. With an empty box, turning it on asks for a prompt first.",
+ "Replacing or clearing all of the enhanced text starts a new prompt, which is enhanced before generating. Small edits still adjust the enhanced version.",
+ "FreeVideo reopens with the last prompt, finished, unfinished or emptied, and its enhancement setting.",
+ "Installing into a folder whose path contains a space, such as \"Comfy UI\", stopped at the PyTorch step with \"File not found\". It installs normally now."
]
},
"zh": {
- "summary": "修复较早安装的用户无法升级到 int8 模型,以及运行环境包无法导入的问题。",
+ "summary": "二次采样新增「先出预览」,作品可以删除,提示词框改进(示例提示词以灰色占位显示,重新打开时恢复上次的提示词),并修复安装路径含空格时无法安装的问题。",
"changes": [
- "较早安装的用户升级到 int8 模型时,会停在「下载比显示的大」。现在显示的下载量包含升级要下载的全部文件(例如采样表),可以正常完成升级。",
- "导入运行环境包时会报「找不到 windows-portable.txt」,现在可以正常导入。"
+ "二次采样新增「先出预览」:先用半分辨率快速出一版预览,满意后点「继续二采」,成片与直接生成完全一致。可以批量出几版预览,在「作品」里挑一版继续二采。",
+ "「作品」里可以删除不需要的视频,删除前会确认。",
+ "示例提示词以灰色占位显示,直接输入即可替换,无需先清空;未输入过时按示例生成,输入后再清空则保持为空。",
+ "提示词增强改写已输入的提示词。提示词框为空时开启增强,会提示先输入提示词。",
+ "在改写版中全选替换或清空后重新输入,会作为新的提示词,生成前自动改写;局部修改仍只调整改写版。",
+ "重新打开 FreeVideo 时恢复上次编写的提示词(包括未写完或已清空的内容)及增强设置。",
+ "安装位置的路径中有空格(例如「Comfy UI」)时,安装会停在 PyTorch 步骤并报「File not found」。现在可以正常安装。"
]
}
}
diff --git a/freevideo_engine/worker.py b/freevideo_engine/worker.py
index 9f119fe..46c3831 100644
--- a/freevideo_engine/worker.py
+++ b/freevideo_engine/worker.py
@@ -205,6 +205,9 @@ def generate(request, resident=None):
from .latent_upscale import upscale
if engine.steps != sampling_plan['base_steps']:
raise ValueError('Engine steps disagree with the requested sampling plan')
+ preview = request.get('preview') is True
+ if preview and (request.get('resume_refine') or request.get('resume_first_pass')):
+ raise ValueError('A preview samples its own first pass')
if request.get('resume_refine'):
# A retry after a second-pass failure: this request's
# retained first pass, latent upscale and crop are
@@ -221,123 +224,168 @@ def generate(request, resident=None):
completed_steps=sampling_plan['base_steps'],
total=sampling_plan['total_steps'])), flush=True)
else:
- pass_budget = (min(request['gpu_budget_bytes'],
- metrics['device_memory'].get('effective_allocator_limit_bytes') or request['gpu_budget_bytes'])
- if request.get('automatic_pass_cache', False) else None)
- first_options = request.get('first_pass_policy', {}).get('profile', {}).get('engine')
- metrics['first_pass_policy'] = request.get('first_pass_policy')
- latents, audio, first = engine.sample(request['conditioning'], request['seed'],
- **sampling_plan['first'], allow_smaller_canvas=True,
- pass_cache_budget_bytes=pass_budget, gpu_reserve_bytes=request.get('gpu_reserve_bytes', 0),
- compute_options=first_options,
- pass_resident_blocks=first_options.get('resident_blocks') if first_options else None,
- step_callback=memory_diagnostics.complete, progress_total=sampling_plan['total_steps'],
- budget_refresh=refresh_budget if live_budget is not None else None)
- metrics.update(sampling_passes=[first], sample_stage='latent_upscale')
- save(request['metrics'], metrics)
- print(json.dumps(dict(event='latent_upscale', completed_steps=sampling_plan['base_steps'],
- total=sampling_plan['total_steps'])), flush=True)
- tick = time.perf_counter()
- if live_budget is not None:
- refresh_budget('upscale')
- need = upscale_workspace(sampling_plan['upscale_target']) if sampling_plan['upscaler_sha256'] else 0
- if need and resident is not None and not resident.ram_fits(2 * 2**30):
- resident.make_room('latent_upscaler', need, ram_need=2 * 2**30)
- torch.cuda.empty_cache()
- def upscale_available():
- free, _ = torch.cuda.mem_get_info()
- budget = metrics['device_memory'].get('effective_allocator_limit_bytes') or request['gpu_budget_bytes']
- return min(free - request.get('gpu_reserve_bytes', 0), budget - torch.cuda.memory_allocated())
- available = upscale_available()
- # Prefer buffer reuse before discarding useful host weights.
- # The estimate selects this path, not permission to run.
- # Actual allocation remains bounded by the worker's CUDA
- # allocator limit. Keep the completed first-pass latents.
- memory_saving = available < need
- torch.cuda.reset_peak_memory_stats()
- lift_canvas = sampling_plan['upscale_target']
- retry_upscale = False
- upscale_failures = []
- metrics['latent_upscale'] = dict(estimated_workspace_bytes=need,
- available_workspace_bytes=max(0, available), attempted_below_estimate=memory_saving,
- buffer_reuse=memory_saving, allocation_retries=0, allocation_failures=upscale_failures)
- save(request['metrics'], metrics)
- try:
- kwargs = dict(memory_saving=True) if memory_saving else {}
- lifted_video, lifted = upscale(latents, request.get('upscaler_checkpoint'),
- lift_canvas['width'], lift_canvas['height'], **kwargs)
- except Exception as error:
- from .adaptive import classify_failure
- if classify_failure(error)['kind'] != 'gpu_oom' or (engine.closed and memory_saving):
- raise
- from .diagnostic_resources import error_details
- upscale_failures.append(error_details(error))
- retry_upscale = True
- if retry_upscale:
- # Leave the failed call's traceback before retrying so
- # its tensors no longer occupy the allocator. Do not
- # sample again or change the temporal normalization.
- if not engine.closed:
- if resident is not None:
- resident.drop('engine', 'Latent upscale allocation needs additional workspace')
- else:
- engine.close()
- if resident is not None:
+ if request.get('resume_first_pass'):
+ # Upscale a finished preview: its retained first pass
+ # replaces sampling it again; upscale and refine follow.
+ from .decode_resume import load_first_pass
+ metrics.update(sample_stage='first_pass_load')
+ save(request['metrics'], metrics)
+ tick = time.perf_counter()
+ latents, audio, first, reused = load_first_pass(request, torch)
+ latents, audio = latents.to('cuda'), audio.to('cuda')
+ metrics.update(reused, first_pass_load_seconds=time.perf_counter() - tick)
+ print(json.dumps(dict(event='first_pass_reused', source='preview',
+ completed_steps=sampling_plan['base_steps'],
+ total=sampling_plan['total_steps'])), flush=True)
+ else:
+ pass_budget = (min(request['gpu_budget_bytes'],
+ metrics['device_memory'].get('effective_allocator_limit_bytes') or request['gpu_budget_bytes'])
+ if request.get('automatic_pass_cache', False) else None)
+ first_options = request.get('first_pass_policy', {}).get('profile', {}).get('engine')
+ metrics['first_pass_policy'] = request.get('first_pass_policy')
+ latents, audio, first = engine.sample(request['conditioning'], request['seed'],
+ **sampling_plan['first'], allow_smaller_canvas=True,
+ pass_cache_budget_bytes=pass_budget, gpu_reserve_bytes=request.get('gpu_reserve_bytes', 0),
+ compute_options=first_options,
+ pass_resident_blocks=first_options.get('resident_blocks') if first_options else None,
+ step_callback=memory_diagnostics.complete,
+ progress_total=sampling_plan['base_steps'] if preview else sampling_plan['total_steps'],
+ budget_refresh=refresh_budget if live_budget is not None else None)
+ metrics['sampling_passes'] = [first]
+ if preview:
+ # Keep the first pass before the latent upscale; an
+ # upscale of this preview starts from these latents.
+ if not artifacts:
+ raise ValueError('A preview needs an artifacts folder for its first pass')
+ if not bool(torch.isfinite(latents).all() and torch.isfinite(audio).all()):
+ raise RuntimeError('Non-finite first-pass latents; refusing to save a corrupted preview')
+ from .decode_resume import refine_provenance, save_first_pass
+ tick = time.perf_counter()
+ provenance = refine_provenance(request)
+ save_first_pass({'video': latents.cpu(), 'audio': audio.cpu(), 'seed': request['seed'],
+ 'geometry': canvas, 'refine_provenance': provenance}, artifacts, torch)
+ metrics.update(preview=True, finite_latents=True, first_pass_checkpoint_complete=True,
+ refine_provenance=provenance, latent_save_seconds=time.perf_counter() - tick)
+ else:
+ metrics['sample_stage'] = 'latent_upscale'
+ save(request['metrics'], metrics)
+ print(json.dumps(dict(event='latent_upscale', completed_steps=sampling_plan['base_steps'],
+ total=sampling_plan['total_steps'])), flush=True)
+ tick = time.perf_counter()
+ if live_budget is not None:
+ refresh_budget('upscale')
+ need = upscale_workspace(sampling_plan['upscale_target']) if sampling_plan['upscaler_sha256'] else 0
+ if need and resident is not None and not resident.ram_fits(2 * 2**30):
resident.make_room('latent_upscaler', need, ram_need=2 * 2**30)
- import gc
- gc.collect()
torch.cuda.empty_cache()
- metrics['latent_upscale'].update(buffer_reuse=True, allocation_retries=1)
+ def upscale_available():
+ free, _ = torch.cuda.mem_get_info()
+ budget = metrics['device_memory'].get('effective_allocator_limit_bytes') or request['gpu_budget_bytes']
+ return min(free - request.get('gpu_reserve_bytes', 0), budget - torch.cuda.memory_allocated())
+ available = upscale_available()
+ # Prefer buffer reuse before discarding useful host weights.
+ # The estimate selects this path, not permission to run.
+ # Actual allocation remains bounded by the worker's CUDA
+ # allocator limit. Keep the completed first-pass latents.
+ memory_saving = available < need
+ torch.cuda.reset_peak_memory_stats()
+ lift_canvas = sampling_plan['upscale_target']
+ retry_upscale = False
+ upscale_failures = []
+ metrics['latent_upscale'] = dict(estimated_workspace_bytes=need,
+ available_workspace_bytes=max(0, available), attempted_below_estimate=memory_saving,
+ buffer_reuse=memory_saving, allocation_retries=0, allocation_failures=upscale_failures)
save(request['metrics'], metrics)
- lifted_video, lifted = upscale(latents, request.get('upscaler_checkpoint'),
- lift_canvas['width'], lift_canvas['height'], memory_saving=True)
- latents = lifted_video
- del lifted_video
- latents = crop_latents(latents, sampling_plan)
- lifted.update(stage_seconds=time.perf_counter() - tick,
- crop=dict(sampling_plan['crop']),
- estimated_workspace_bytes=need,
- available_workspace_bytes=max(0, available),
- attempted_below_estimate=memory_saving,
- allocation_retries=len(upscale_failures),
- allocation_failures=upscale_failures,
- torch_peak_allocated_bytes=torch.cuda.max_memory_allocated(),
- torch_peak_reserved_bytes=torch.cuda.max_memory_reserved(),
- transformer_reloaded=engine.closed)
- metrics['latent_upscale'] = lifted
- if engine.closed:
- tick = time.perf_counter()
- engine, _ = resident.engine(request, factory) if resident is not None else (factory(), False)
- reload_seconds = time.perf_counter() - tick
- metrics['load_seconds'] += reload_seconds
- metrics['load_breakdown']['two_pass_reload_seconds'] = reload_seconds
- if artifacts:
- # Retain the second pass's input so an out-of-memory
- # retry starts there. A failed save only loses that.
try:
- from .decode_resume import refine_provenance, save_refine_input
- provenance = refine_provenance(request)
- save_refine_input({'video': latents.cpu(), 'audio': audio.cpu(), 'seed': request['seed'],
- 'geometry': canvas, 'refine_provenance': provenance}, artifacts, torch)
- metrics.update(refine_checkpoint_complete=True, refine_provenance=provenance,
- refine_source_attempt=request.get('resource_attempt'))
- except (OSError, RuntimeError, ValueError) as error:
- metrics['refine_checkpoint_error'] = str(error)
- metrics['sample_stage'] = 'refine'
- save(request['metrics'], metrics)
- def refined_complete(video, sound, receipt):
- sampling_complete(video, sound, combine(first, receipt, lifted, sampling_plan))
- restart_seed = (request['seed'] + sampling_plan['restart_seed_offset']) % (1 << 64)
- latents, audio, second = engine.sample(request['conditioning'], restart_seed,
- **sampling_plan['second'], initial_latents=(latents, audio), refine_steps=sampling_plan['refine_steps'],
- refine_schedule=sampling_plan.get('refine_schedule'),
- final_step_callback=None if decoder_cached else decode_read_ahead.start,
- step_callback=memory_diagnostics.complete, sampling_complete_callback=refined_complete,
- progress_offset=sampling_plan['base_steps'], progress_total=sampling_plan['total_steps'],
- gpu_reserve_bytes=request.get('gpu_reserve_bytes', 0),
- budget_refresh=refresh_budget if live_budget is not None else None)
- sampled = combine(first, second, lifted, sampling_plan)
- metrics.pop('sample_stage', None)
+ kwargs = dict(memory_saving=True) if memory_saving else {}
+ lifted_video, lifted = upscale(latents, request.get('upscaler_checkpoint'),
+ lift_canvas['width'], lift_canvas['height'], **kwargs)
+ except Exception as error:
+ from .adaptive import classify_failure
+ if classify_failure(error)['kind'] != 'gpu_oom' or (engine.closed and memory_saving):
+ raise
+ from .diagnostic_resources import error_details
+ upscale_failures.append(error_details(error))
+ retry_upscale = True
+ if retry_upscale:
+ # Leave the failed call's traceback before retrying so
+ # its tensors no longer occupy the allocator. Do not
+ # sample again or change the temporal normalization.
+ if not engine.closed:
+ if resident is not None:
+ resident.drop('engine', 'Latent upscale allocation needs additional workspace')
+ else:
+ engine.close()
+ if resident is not None:
+ resident.make_room('latent_upscaler', need, ram_need=2 * 2**30)
+ import gc
+ gc.collect()
+ torch.cuda.empty_cache()
+ metrics['latent_upscale'].update(buffer_reuse=True, allocation_retries=1)
+ save(request['metrics'], metrics)
+ lifted_video, lifted = upscale(latents, request.get('upscaler_checkpoint'),
+ lift_canvas['width'], lift_canvas['height'], memory_saving=True)
+ latents = lifted_video
+ del lifted_video
+ latents = crop_latents(latents, sampling_plan)
+ lifted.update(stage_seconds=time.perf_counter() - tick,
+ crop=dict(sampling_plan['crop']),
+ estimated_workspace_bytes=need,
+ available_workspace_bytes=max(0, available),
+ attempted_below_estimate=memory_saving,
+ allocation_retries=len(upscale_failures),
+ allocation_failures=upscale_failures,
+ torch_peak_allocated_bytes=torch.cuda.max_memory_allocated(),
+ torch_peak_reserved_bytes=torch.cuda.max_memory_reserved(),
+ transformer_reloaded=engine.closed)
+ metrics['latent_upscale'] = lifted
+ if engine.closed:
+ tick = time.perf_counter()
+ engine, _ = resident.engine(request, factory) if resident is not None else (factory(), False)
+ reload_seconds = time.perf_counter() - tick
+ metrics['load_seconds'] += reload_seconds
+ metrics['load_breakdown']['two_pass_reload_seconds'] = reload_seconds
+ if artifacts:
+ # Retain the second pass's input so an out-of-memory
+ # retry starts there. A failed save only loses that.
+ try:
+ from .decode_resume import refine_provenance, save_refine_input
+ provenance = refine_provenance(request)
+ save_refine_input({'video': latents.cpu(), 'audio': audio.cpu(), 'seed': request['seed'],
+ 'geometry': canvas, 'refine_provenance': provenance}, artifacts, torch)
+ metrics.update(refine_checkpoint_complete=True, refine_provenance=provenance,
+ refine_source_attempt=request.get('resource_attempt'))
+ except (OSError, RuntimeError, ValueError) as error:
+ metrics['refine_checkpoint_error'] = str(error)
+ if preview:
+ sampled = dict(first, sampling_plan=sampling_plan, sampling_passes=[dict(first)],
+ step_seconds=list(first['step_seconds']), sample_seconds=first['sample_seconds'])
+ if not decoder_cached:
+ decode_read_ahead.start()
+ else:
+ metrics['sample_stage'] = 'refine'
+ save(request['metrics'], metrics)
+ def refined(receipt):
+ result = combine(first, receipt, lifted, sampling_plan)
+ if request.get('resume_first_pass'):
+ # This request sampled only the refinement; the
+ # reused first pass's seconds belong to its preview.
+ result.update(reused_first_pass_seconds=first['sample_seconds'],
+ sample_seconds=result['sample_seconds'] - first['sample_seconds'])
+ return result
+ def refined_complete(video, sound, receipt):
+ sampling_complete(video, sound, refined(receipt))
+ restart_seed = (request['seed'] + sampling_plan['restart_seed_offset']) % (1 << 64)
+ latents, audio, second = engine.sample(request['conditioning'], restart_seed,
+ **sampling_plan['second'], initial_latents=(latents, audio), refine_steps=sampling_plan['refine_steps'],
+ refine_schedule=sampling_plan.get('refine_schedule'),
+ final_step_callback=None if decoder_cached else decode_read_ahead.start,
+ step_callback=memory_diagnostics.complete, sampling_complete_callback=refined_complete,
+ progress_offset=sampling_plan['base_steps'], progress_total=sampling_plan['total_steps'],
+ gpu_reserve_bytes=request.get('gpu_reserve_bytes', 0),
+ budget_refresh=refresh_budget if live_budget is not None else None)
+ sampled = refined(second)
+ metrics.pop('sample_stage', None)
else:
latents, audio, sampled = engine.sample(request['conditioning'], request['seed'],
frames=canvas.get('frames', 243), width=canvas.get('width', 1344), height=canvas.get('height', 768),
diff --git a/scripts/test_freevideo_entry.mjs b/scripts/test_freevideo_entry.mjs
index 8cce8f3..410c0ad 100644
--- a/scripts/test_freevideo_entry.mjs
+++ b/scripts/test_freevideo_entry.mjs
@@ -2,7 +2,7 @@ import assert from 'node:assert/strict';
import {readFile} from 'node:fs/promises';
import test from 'node:test';
import {outputDownloadURL} from '../web/output_download.js';
-import {regenerateResult} from '../web/studio_queue.js';
+import {regenerateResult, upscaleResult} from '../web/studio_queue.js';
class Element {
constructor(tag) {
@@ -29,7 +29,7 @@ test('main browser entry registers both views and preserves node hooks', async (
const api = Object.assign(events, {apiURL: value => value});
const document = {head: new Element('head'), createElement: tag => new Element(tag)};
const dependencies = {
- app, api, outputDownloadURL, regenerateResult,
+ app, api, outputDownloadURL, regenerateResult, upscaleResult,
createErrorPanel: () => ({element: new Element('error'), show() {}, clear() {}}),
errorText: value => String(value),
openStudio: node => opened.push(node),
@@ -43,6 +43,7 @@ test('main browser entry registers both views and preserves node hooks', async (
refreshNavigation: () => refreshed++, preferredView: () => null,
attachReferencePicker() {}, referenceItems: () => [], syncReferencePrompt() {},
shareButton: () => new Element('button'),
+ rememberPromptDraft() {}, savePromptDraft: async () => false,
};
const previous = new Map();
for (const [name, value] of Object.entries({
@@ -67,7 +68,8 @@ test('main browser entry registers both views and preserves node hooks', async (
.replace("import { attachReferencePicker, referenceItems, syncReferencePrompt } from './prompt_references.js';", 'const {attachReferencePicker,referenceItems,syncReferencePrompt} = globalThis.__freevideoEntryTest;')
.replace("import { outputDownloadURL } from './output_download.js';", 'const {outputDownloadURL} = globalThis.__freevideoEntryTest;')
.replace("import { shareButton } from './share.js';", 'const {shareButton} = globalThis.__freevideoEntryTest;')
- .replace("import { regenerateResult } from './studio_queue.js';", 'const {regenerateResult} = globalThis.__freevideoEntryTest;')
+ .replace("import { rememberPromptDraft, savePromptDraft } from './prompt_draft.js';", 'const {rememberPromptDraft,savePromptDraft} = globalThis.__freevideoEntryTest;')
+ .replace("import { regenerateResult, upscaleResult } from './studio_queue.js';", 'const {regenerateResult,upscaleResult} = globalThis.__freevideoEntryTest;')
.replace("import { installNavigation, refreshNavigation, preferredView } from './view_navigation.js';", 'const {installNavigation,refreshNavigation,preferredView} = globalThis.__freevideoEntryTest;');
await import('data:text/javascript;base64,' + Buffer.from(source).toString('base64'));
const extension = extensions.find(row => row.name === 'FreeVideo.UnifiedMedia');
diff --git a/web/freevideo.js b/web/freevideo.js
index 9d16587..f47b7af 100644
--- a/web/freevideo.js
+++ b/web/freevideo.js
@@ -10,7 +10,8 @@ import { installNavigation, refreshNavigation, preferredView } from './view_navi
import { attachReferencePicker, referenceItems, syncReferencePrompt } from './prompt_references.js';
import { outputDownloadURL } from './output_download.js';
import { shareButton } from './share.js';
-import { regenerateResult } from './studio_queue.js';
+import { rememberPromptDraft, savePromptDraft } from './prompt_draft.js';
+import { regenerateResult, upscaleResult } from './studio_queue.js';
const languageOverride = typeof location !== 'undefined'
? new URLSearchParams(location.search).get('freevideo_lang') : null;
@@ -280,9 +281,13 @@ function resultPanel(node) {
if (!note) { note = el('div', '', 'fv-note fv-prewarm'); panel.append(note); }
const onGPU = Number.isFinite(value.gpu_bytes_after);
const size = onGPU ? ` · ${(value.gpu_bytes_after / 2 ** 30).toFixed(1)} GiB` : '';
- note.textContent = value.state === 'loading' ? text('Preparing text encoder on GPU in background', '正在后台预加载文本编码器到 GPU') + size
+ const percent = ` · ${Math.min(100, Math.floor(100 * value.done_bytes / value.bytes))}%`;
+ note.textContent = value.purpose === 'upscale'
+ ? (value.state === 'reading' ? text('Preparing the files the second pass reads', '正在预读二采要用的模型文件') + percent
+ : value.state === 'ready' ? text('Second-pass files warmed · no VRAM reserved', '二采要用的模型文件已预热 · 不占用显存') : '')
+ : value.state === 'loading' ? text('Preparing text encoder on GPU in background', '正在后台预加载文本编码器到 GPU') + size
: value.state === 'partial' ? text('Text encoder partially prepared on GPU', '文本编码器已部分预加载到 GPU') + size
- : value.state === 'reading' ? text('Preparing encoder files', '正在预读编码器文件') + ` · ${Math.min(100, Math.floor(100 * value.done_bytes / value.bytes))}%`
+ : value.state === 'reading' ? text('Preparing encoder files', '正在预读编码器文件') + percent
: value.state === 'ready' ? (onGPU ? text('Text encoder ready on GPU', '文本编码器已在 GPU 就绪') + size
: text('Encoder files warmed · no VRAM reserved', '编码器文件已预热 · 不占用显存')) : '';
node.freevideoPrewarm = note.textContent;
@@ -309,6 +314,16 @@ function resultPanel(node) {
}
links.append(shareButton(value, text));
links.append(el("span", value.result_cache_hit ? text("Reused previous result", "已复用上次结果") : value.conditioning_cache_hit ? text("Input cache reused", "已复用输入缓存") : text("Inputs encoded", "已编码输入"), "fv-mode"));
+ if (value.preview) {
+ const upscale = el('button', text('Run second pass', '继续二采')); upscale.type = 'button';
+ upscale.title = text('Continue from this preview: latent upscale, the second pass and a full-resolution decode.', '从这个预览接着跑:潜空间放大、二采,再以全分辨率解码。');
+ upscale.onclick = async () => {
+ upscale.disabled = true;
+ try { await node.freevideoUpscaleResult(value); upscale.textContent = text('Queued', '已加入队列'); }
+ catch (error) { upscale.disabled = false; node.freevideoShowFailure({exception_message: error.message}); }
+ };
+ links.append(upscale);
+ }
if (value.result_cache_hit) {
const again = el('button', text('Regenerate', '重新生成')); again.type = 'button';
again.onclick = async () => {
@@ -325,6 +340,12 @@ function resultPanel(node) {
? text("Unified memory: device capacity. Process RAM is not total GPU memory, ", "统一内存为设备总量;进程内存不代表 GPU 内存总占用,")
: text("VRAM: engine allocator peak. RAM: measured request processes, ", "显存为引擎分配器峰值,内存为请求进程实测,")) + (value.ram_metric || "unknown");
};
+ node.freevideoUpscaleResult = async value => {
+ if (node.freevideoUpscaling) return;
+ node.freevideoUpscaling = true;
+ try { return await upscaleResult(api, node.id, value, () => app.graphToPrompt()); }
+ finally { node.freevideoUpscaling = false; }
+ };
node.freevideoRegenerateResult = async value => {
if (node.freevideoRegenerating) return;
node.freevideoRegenerating = true;
@@ -359,7 +380,11 @@ const progressConnection = createProgressConnection(api, message => {
app.registerExtension({
name: "FreeVideo.UnifiedMedia",
beforeConfigureGraph() { configuringGraph = true; },
- async setup() { progressConnection.start(); installNavigation(openStudio); startUpdateChecks(); await notifyCompatibility(); },
+ async setup() {
+ progressConnection.start(); installNavigation(openStudio); startUpdateChecks();
+ window.addEventListener('pagehide', () => savePromptDraft(api, true));
+ await notifyCompatibility();
+ },
async beforeRegisterNodeDef(nodeType, nodeData) {
if (nodeData.name === 'FreeVideoReference') {
const connected = nodeType.prototype.onConnectionsChange;
@@ -367,6 +392,7 @@ app.registerExtension({
return;
}
if (!["FreeVideoMedia", "FreeVideoGenerate", "FreeVideoLoRAStack"].includes(nodeData.name)) return;
+ if (nodeData.name === "FreeVideoGenerate") nodeType.prototype.freevideoStarterPrompt = nodeData.input?.required?.text?.[1]?.default || '';
const created = nodeType.prototype.onNodeCreated;
nodeType.prototype.onNodeCreated = function () {
const result = created?.apply(this, arguments);
@@ -401,6 +427,13 @@ app.registerExtension({
'修改采样步数可能降低生成质量。默认一采 8 步、二采 3 步;其他二采步数需小于一采步数。');
}
}
+ const prompt = this.widgets?.find(w => w.name === 'text');
+ if (prompt) {
+ // Every edit, in the creation panel or on the node, is the draft the launcher reopens.
+ const node = this, callback = prompt.callback;
+ prompt.callback = function (...args) { if (!configuringGraph) rememberPromptDraft(api, node); return callback?.apply(this, args); };
+ prompt.inputEl?.addEventListener?.('input', () => rememberPromptDraft(api, node));
+ }
resultPanel(this);
this.addWidget('button', text('Open creative workspace', '打开创作面板'), null, () => openStudio(this), {serialize: false});
const removed = this.onRemoved;
diff --git a/web/generation_progress.js b/web/generation_progress.js
index 30dcff1..c9d163d 100644
--- a/web/generation_progress.js
+++ b/web/generation_progress.js
@@ -41,6 +41,21 @@ function encodingLabel(label, t) {
return t(...(labels[label] || ['Preparing prompt and references', '准备提示词与参考素材']));
}
+// The bridge describes the sampling plan in English; say it in the UI language.
+function planText(message, t) {
+ const plan = message.sampling_plan, first = plan?.first, second = plan?.second;
+ if (!plan || !first) return null;
+ const base = plan.base_steps ?? 8, refine = plan.refine_steps ?? 3;
+ if (plan.enabled && plan.preview) return {
+ label: t(`First-pass preview · ${first.width} × ${first.height} · ${base} steps`, `一采预览 · ${first.width} × ${first.height} · ${base} 步`),
+ detail: t('Run the second pass from the result when it looks right', '满意后在结果中继续二采')};
+ if (plan.enabled && second) return {
+ label: t(`Two-pass · ${first.width} × ${first.height} → ${second.width} × ${second.height} · ${base} + ${refine} steps`,
+ `二次采样 · ${first.width} × ${first.height} → ${second.width} × ${second.height} · ${base} + ${refine} 步`),
+ detail: t('Audio is preserved from the first pass', '音频沿用一采结果')};
+ return {label: t(`Single-pass · ${base} steps`, `单次采样 · ${base} 步`), detail: message.detail || ''};
+}
+
// A reference clip longer than the video being generated is cut to the same
// length, at most 15 s, as in the official pipeline. Say so; never cut silently.
export function referenceTrimText(row, t) {
@@ -314,7 +329,9 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
transformer_release: t('Preparing video decoding', '正在准备视频解码'),
}[message.stage] || t('Preparing video decoding', '正在准备视频解码')
: message.timing_phase === 'encoding' ? encodingLabel(message.label, t)
- : message.label || t('Preparing video', '正在准备视频');
+ : message.stage === 'preview_mismatch'
+ ? t('The preview no longer matches; generating the whole video', '预览的中间结果已对不上,改为完整生成')
+ : planText(message, t)?.label || message.label || t('Preparing video', '正在准备视频');
if (compact && !complete) {
const phase = message.phase || message.timing_phase;
label.textContent = sampling ? `${t('Sampling', '采样')} ${message.done} / ${message.total}`
@@ -338,7 +355,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
: message.stage === 'inputs' ? t('Preparing sampling inputs', '正在准备采样输入')
: message.done > 0 ? t(`${message.done} steps complete`, `已完成 ${message.done} 步`)
: t('Starting the first step', '正在开始第一步');
- } else detail.textContent = message.detail || '';
+ } else detail.textContent = planText(message, t)?.detail ?? message.detail ?? '';
if ((downloadLabel || message.phase === 'dependencies') && valid(message.done) && valid(message.total)) {
detail.textContent = `${(message.done/2**20).toFixed(1)} / ${(message.total/2**20).toFixed(1)} MiB`
+ (valid(message.bytes_per_second) && message.bytes_per_second > 0 ? ` · ${(message.bytes_per_second/2**20).toFixed(1)} MiB/s` : '');
diff --git a/web/launcher.js b/web/launcher.js
index 7ce70cd..2b6490a 100644
--- a/web/launcher.js
+++ b/web/launcher.js
@@ -1,5 +1,6 @@
import { app } from '../../scripts/app.js';
import { api } from '../../scripts/api.js';
+import { restorePromptDraft } from './prompt_draft.js';
// Explicit launcher navigation only; ordinary ComfyUI visits restore their own
// graph. Remove the flag before loading so refreshing never resets user edits.
@@ -13,7 +14,8 @@ app.registerExtension({
try {
const response = await api.fetchApi('/freevideo/launcher/workflow');
if (!response.ok) throw new Error(`FreeVideo workflow: HTTP ${response.status}`);
- const workflow = await response.json();
+ // A fresh copy of the bundled workflow, with the prompt last written.
+ const workflow = await restorePromptDraft(api, await response.json());
await app.loadGraphData(workflow, true, true, 'FreeVideo');
} catch (error) {
app.extensionManager?.toast?.add({severity: 'error', summary: 'FreeVideo', detail: String(error), life: 15000});
diff --git a/web/library.css b/web/library.css
index 27d0066..7f8f21c 100644
--- a/web/library.css
+++ b/web/library.css
@@ -34,3 +34,12 @@
.fv-library .fv-backfill p{margin:0;font-size:12px;line-height:1.6;color:var(--fv-muted)}
.fv-library .fv-backfill-actions{display:flex;justify-content:center;gap:8px;margin-top:6px}
.fv-library .fv-backfill-actions button{min-height:30px;padding:4px 12px;font-size:12px}
+.fv-library-picture{position:relative}
+/* The picture, its caption and actions form one block, centred in the detail. */
+.fv-library-detail{container-type:inline-size}.fv-library-detail::after{content:'';margin-top:auto}
+.fv-library-frame{position:relative;container-type:size;max-height:max(160px,calc(100cqw / var(--fv-ratio,1.7778)));margin-top:auto;background:none;box-shadow:none}
+.fv-library-media{position:relative;width:min(100cqw,calc(100cqh * var(--fv-ratio,1.7778)));aspect-ratio:var(--fv-ratio,1.7778);max-width:100%;max-height:100%;overflow:hidden;border-radius:var(--fv-r-md);background:#0d1118;box-shadow:0 12px 40px #0003}
+.fv-library-media video{width:100%;height:100%}
+.fv-library .fv-library-picture .fv-preview-badge{top:6px;left:6px;padding:1px 8px;font-size:10px}
+.fv-library .fv-library-notice{min-height:0;margin:6px 0 0;text-align:center}
+.fv-library .fv-library-notice:empty{display:none}
diff --git a/web/library.js b/web/library.js
index cadff90..f208916 100644
--- a/web/library.js
+++ b/web/library.js
@@ -1,6 +1,6 @@
import { api } from '../../scripts/api.js';
import { closeDialog } from './motion.js';
-import { resultActions } from './result_actions.js';
+import { resultActions, previewBadge } from './result_actions.js';
const style = document.createElement('link');
style.rel = 'stylesheet'; style.href = new URL('./library.css', import.meta.url).href; document.head.append(style);
@@ -54,7 +54,8 @@ async function workflowBackfill(collection, t) {
actions.append(run, later);
}
-export function openLibrary(t) {
+// ``upscale`` queues the second pass of a first-pass preview (the studio's node).
+export function openLibrary(t, {upscale = null} = {}) {
if (opened?.open) { opened.focus(); return; }
const dialog = el('dialog', null, 'fv-studio fv-library'); opened = dialog;
dialog.setAttribute('aria-label', t('Your creations', '我的作品'));
@@ -81,7 +82,8 @@ export function openLibrary(t) {
stats.setAttribute('aria-label', t('Generation statistics', '生成统计'));
const budget = el('div', '', 'fv-budget');
const links = el('div', null, 'fv-result-links');
- detail.append(all, frame, caption, links, stats, budget);
+ const notice = el('p', '', 'fv-library-message fv-library-notice'); notice.setAttribute('role', 'status');
+ detail.append(all, frame, caption, links, notice, stats, budget);
body.append(collection, detail); dialog.append(header, body);
let disposed = false, loading = false, next = null, selected = null, player = null, arrivals = [];
const rows = new Map(), cards = new Map();
@@ -94,7 +96,9 @@ export function openLibrary(t) {
};
const dimensions = row => {
const g = row.geometry || {}, parts = [];
- if (g.width && g.height) parts.push(`${g.width} × ${g.height}`);
+ // A preview's file is its first pass; geometry names the finished size.
+ const shown = row.preview && row.preview_geometry?.width ? row.preview_geometry : g;
+ if (shown.width && shown.height) parts.push(`${shown.width} × ${shown.height}`);
const duration = g.seconds || (g.frames && g.frames / (g.fps || 24));
if (duration) parts.push(`${Number(duration).toFixed(1)} s`);
return parts.join(' · ');
@@ -103,20 +107,33 @@ export function openLibrary(t) {
if (!player) return;
player.pause(); player.removeAttribute('src'); player.load(); player = null;
}
+ function mountPlayer(row) {
+ releasePlayer(); frame.replaceChildren();
+ player = el('video'); player.controls = true; player.preload = 'metadata'; player.playsInline = true;
+ player.src = view(row.video);
+ // The picture's own box, so a preview's badge sits on its corner.
+ const media = el('div', null, 'fv-library-media');
+ const g = row.geometry || {};
+ if (g.width > 0 && g.height > 0) frame.style.setProperty('--fv-ratio', String(g.width / g.height));
+ media.append(player);
+ if (row.preview) media.append(previewBadge(row, t));
+ frame.append(media);
+ player.addEventListener('error', () => {
+ if (!disposed && selected === row.video) {
+ const note = el('p', t('This video is unavailable. Refresh if it was moved.', '视频暂时无法打开,若文件已移动请刷新。'), 'fv-library-message');
+ frame.replaceChildren(note);
+ }
+ }, {once: true});
+ }
function select(row, reveal = true) {
if (selected !== row.video) {
- releasePlayer(); frame.replaceChildren();
selected = row.video;
- player = el('video'); player.controls = true; player.preload = 'metadata'; player.playsInline = true;
- player.src = view(row.video); frame.append(player);
- player.addEventListener('error', () => {
- if (!disposed && selected === row.video) {
- const note = el('p', t('This video is unavailable. Refresh if it was moved.', '视频暂时无法打开,若文件已移动请刷新。'), 'fv-library-message');
- frame.replaceChildren(note);
- }
- }, {once: true});
+ mountPlayer(row);
date.textContent = timestamp(row); geometry.textContent = dimensions(row);
- links.replaceChildren(resultActions(row, t));
+ links.replaceChildren(resultActions(row, t, {
+ secondPass: row.preview && upscale ? () => secondPass(row) : null,
+ remove: () => remove(row)}));
+ notice.textContent = '';
for (const [file, card] of cards) card.setAttribute('aria-pressed', String(file === selected));
}
// These are the saved generation's measurements, including when the
@@ -133,11 +150,51 @@ export function openLibrary(t) {
const item = el('div', null, 'fv-stat');
item.append(el('strong', number(value, scale, unit)), el('span', label)); stats.append(item);
}
- stats.hidden = ![row.sample_seconds, row.request_seconds, row.vram_peak_bytes, row.ram_peak_bytes].some(measured);
- budget.textContent = measured(row.gpu_budget_bytes)
+ // A preview's timings describe only its first pass; they stay in its report.
+ stats.hidden = row.preview || ![row.sample_seconds, row.request_seconds, row.vram_peak_bytes, row.ram_peak_bytes].some(measured);
+ budget.textContent = !row.preview && measured(row.gpu_budget_bytes)
? `${t('VRAM budget', '可用显存预算')} ${number(row.gpu_budget_bytes, 2 ** 30, 'GiB')} · ${t('Device', '显卡总量')} ${number(row.gpu_total_bytes, 2 ** 30, 'GiB')}` : '';
if (reveal) dialog.dataset.detail = 'true';
}
+ async function secondPass(row) {
+ notice.textContent = '';
+ try {
+ await upscale(row);
+ notice.textContent = t('Queued. The finished video appears here when the second pass is done.', '已加入队列,二采完成后成片会出现在作品里。');
+ } catch (error) {
+ notice.textContent = t('Could not start the second pass: ', '无法开始二采:') + error.message;
+ throw error;
+ }
+ }
+ // Throws a readable Error for the confirmation to show; on success the
+ // creation leaves the collection.
+ async function remove(row) {
+ // Windows keeps a streamed video open; release it before deleting.
+ if (selected === row.video) releasePlayer();
+ let value = {};
+ try {
+ const response = await api.fetchApi('/freevideo/library/delete', {method: 'POST',
+ headers: {'Content-Type': 'application/json'}, body: JSON.stringify({id: row.id})});
+ value = await response.json().catch(() => ({}));
+ if (!response.ok || !value.deleted) throw new Error(value.reason || 'failed');
+ } catch (error) {
+ if (!disposed && selected === row.video) mountPlayer(row);
+ throw new Error(error.message === 'foreign-files'
+ ? t('Not deleted: the folder also holds files FreeVideo did not write: ', '未删除:文件夹里还有不是 FreeVideo 生成的文件:') + (value.kept || []).join(', ')
+ : error.message === 'in-use' ? t('Not deleted: its second pass is running, or the video is still open. Try again in a moment.', '未删除:正在用它跑二采,或视频仍被占用,请稍后再试。')
+ : t('Could not delete this creation.', '无法删除这条作品。'));
+ }
+ if (disposed) return;
+ rows.delete(row.video); cards.get(row.video)?.remove(); cards.delete(row.video);
+ if (selected === row.video) {
+ selected = null; frame.replaceChildren(); links.replaceChildren(); date.textContent = geometry.textContent = '';
+ stats.hidden = true; stats.replaceChildren(); budget.textContent = '';
+ const next = rows.values().next().value;
+ if (next && !narrow.matches) select(next, false); else dialog.dataset.detail = 'false';
+ }
+ notice.textContent = t('Deleted.', '已删除。');
+ if (!rows.size) message.textContent = t('Your finished videos will appear here.', '生成完成的视频会保存在这里。');
+ }
const thumbnails = new IntersectionObserver(entries => {
for (const {target, isIntersecting} of entries) if (isIntersecting) {
target.src = target.dataset.src; thumbnails.unobserve(target);
@@ -155,7 +212,9 @@ export function openLibrary(t) {
const id = row.id || row.video.replace(/^FreeVideo\//, '').replace(/\/video\.mp4$/, '');
image.dataset.src = api.apiURL('/freevideo/library/thumbnail?' + new URLSearchParams({id}));
image.onerror = () => { image.hidden = true; };
- picture.append(image); card.append(picture, el('span', timestamp(row)), el('small', dimensions(row)));
+ picture.append(image);
+ if (row.preview) picture.append(el('span', t('First-pass preview', '一采预览'), 'fv-preview-badge'));
+ card.append(picture, el('span', timestamp(row)), el('small', dimensions(row)));
cards.set(row.video, card); first ? grid.prepend(card) : grid.append(card); thumbnails.observe(image);
}
async function load(appendPage) {
diff --git a/web/prompt_draft.js b/web/prompt_draft.js
new file mode 100644
index 0000000..28ccdb7
--- /dev/null
+++ b/web/prompt_draft.js
@@ -0,0 +1,48 @@
+// The launcher opens a fresh copy of the bundled workflow each time, so the
+// prompt being written, finished or not, is kept in ComfyUI's user data and put
+// back into that copy. The untouched starter prompt is the default, not a draft;
+// a box emptied after typing is one, and reopens empty.
+const FILE = '/userdata/' + encodeURIComponent('freevideo/prompt-draft.json');
+let timer = null, pending = null;
+
+export function promptDraft(node) {
+ const text = node?.widgets?.find(w => w.name === 'text')?.value;
+ if (typeof text !== 'string') return null;
+ const versions = node.properties?.freevideo_prompt_versions;
+ return {schema: 1, text: node.freevideoStarterPrompt && text === node.freevideoStarterPrompt ? null : text,
+ versions: versions && typeof versions === 'object' ? versions : null};
+}
+
+export function applyPromptDraft(workflow, draft) {
+ const node = workflow?.nodes?.find(row => row.type === 'FreeVideoGenerate');
+ if (!node || !Array.isArray(node.widgets_values) || draft?.schema !== 1) return workflow;
+ if (typeof draft.text === 'string') node.widgets_values[0] = draft.text;
+ const versions = draft.versions;
+ if (versions && typeof versions.original === 'string' && typeof versions.rewritten === 'string')
+ node.properties = {...node.properties, freevideo_prompt_versions: versions};
+ return workflow;
+}
+
+export function rememberPromptDraft(api, node) {
+ pending = node; clearTimeout(timer);
+ timer = setTimeout(() => savePromptDraft(api), 500);
+}
+
+// keepalive lets the last edit outlive a closing page.
+export async function savePromptDraft(api, keepalive = false) {
+ clearTimeout(timer); timer = null;
+ const draft = promptDraft(pending); pending = null;
+ if (!draft) return false;
+ try {
+ const reply = await api.fetchApi(FILE + '?overwrite=true', {method: 'POST', body: JSON.stringify(draft), keepalive});
+ return reply.ok;
+ } catch { return false; }
+}
+
+export async function restorePromptDraft(api, workflow) {
+ try {
+ const reply = await api.fetchApi(FILE, {cache: 'no-store'});
+ if (reply.ok) return applyPromptDraft(workflow, await reply.json());
+ } catch { /* No draft yet, or unreadable: open the bundled prompt. */ }
+ return workflow;
+}
diff --git a/web/prompt_enhance.js b/web/prompt_enhance.js
index 7eb5f8c..69841f9 100644
--- a/web/prompt_enhance.js
+++ b/web/prompt_enhance.js
@@ -7,8 +7,14 @@ export class PromptVersions {
if (this.text() !== text) this.edit(text);
}
text() { return this.value[this.value.selected]; }
- edit(text) {
+ // Clearing or replacing all of the enhanced text starts a new scene rather than
+ // a tweak of the rewrite, so a later rewrite starts from it, not the old original.
+ edit(text, restart = false) {
this.value.edits = (this.value.edits || 0) + 1;
+ if (restart && this.value.selected === 'rewritten') {
+ Object.assign(this.value, {original: text, rewritten: '', selected: 'original', context: '', useOriginal: false});
+ return;
+ }
this.value[this.value.selected] = text;
if (this.value.selected === 'original') this.value.context = '';
}
@@ -22,8 +28,12 @@ export class PromptVersions {
select(name) { if (name === 'original' || this.value.rewritten) { this.value.selected = name; this.value.useOriginal = name === 'original'; } }
}
-export function createPromptEnhancer({node, input, editor, api, context, setText, t, onReport = () => {}}) {
- const versions = new PromptVersions(node.properties?.freevideo_prompt_versions, input.value);
+// An empty box is never rewritten, the untouched starter prompt included:
+// enhancement waits for the user's own scene.
+export function createPromptEnhancer({node, input, editor, api, context, setText, t, starter = '', onReport = () => {}, onChange = () => {}}) {
+ const saved = node.properties?.freevideo_prompt_versions;
+ // Drafts saved before the starter became a placeholder hold it as the original.
+ const versions = new PromptVersions(starter && saved?.original === starter ? {...saved, original: ''} : saved, input.value);
const el = (tag, text, cls) => { const e = document.createElement(tag); if (text) e.textContent = text; if (cls) e.className = cls; return e; };
const element = el('div', null, 'fv-enhance');
const bar = el('div', null, 'fv-enhance-bar'), label = el('label', null, 'fv-enhance-toggle');
@@ -63,9 +73,10 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
const actions = el('div', null, 'fv-enhance-install-actions'); actions.append(agree, decline, license);
install.append(detail, actions);
bar.append(label, tabs, redo, cancel); element.append(bar, editor || input, status, install);
- let disposed = false, token = '', job = '', running = null, consent = null, cancelling = false;
+ let disposed = false, token = '', job = '', running = null, consent = null, cancelling = false, waiting = false, replacing = false;
const errorText = error => ({
busy: t('Finish the current queue before enhancing this prompt.', '请等待当前队列完成后再改写。'),
+ prompt_required: t('Type a prompt first; enhancement rewrites it before generating.', '请先输入提示词,生成前会自动改写。'),
setup_required: t('Complete FreeVideo installation first.', '请先完成 FreeVideo 安装。'),
model_missing: t('Download the optional model to enable enhancement.', '下载可选模型后即可启用。'),
download_failed: t('Download paused. Retry to resume.', '下载已暂停,重试即可继续。'),
@@ -82,10 +93,11 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
cancelled: t('Cancelled. Your original prompt is retained.', '已取消,原文已保留。'),
worker_cleanup_failed: t('The rewrite process could not close. Restart FreeVideo before generating.', '改写进程未能退出,请重启 FreeVideo 后再生成。'),
}[error?.message] || t('Could not enhance the prompt. Retry or switch enhancement off to use the original.', '暂时无法改写,请重试,或关闭增强使用原文。'));
- function save() {
+ function save(changed = true) {
node.properties ||= {};
node.properties.freevideo_prompt_versions = {...versions.value};
node.graph?.change();
+ if (changed) onChange();
}
function render() {
toggle.checked = Boolean(versions.value.enabled);
@@ -102,7 +114,12 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
element.dataset.enabled = String(Boolean(versions.value.enabled));
}
function apply() { input.value = versions.text(); setText(input.value); save(); render(); }
- function changed() { versions.edit(input.value); save(); render(); }
+ function changed() {
+ if (waiting) { waiting = false; show(''); }
+ versions.edit(input.value, replacing || !input.value.trim()); replacing = false; save(); render();
+ }
+ const replace = () => { replacing = input.value.length > 0 && input.selectionStart === 0 && input.selectionEnd === input.value.length; };
+ input.addEventListener('beforeinput', replace);
input.addEventListener('input', changed);
async function post(action, body) {
const reply = await api.fetchApi('/freevideo/prompt-vlm/' + action, {method: 'POST',
@@ -146,6 +163,10 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
}
async function enhance() {
if (running) return running;
+ if (!versions.value.original.trim()) {
+ waiting = true; show(errorText({message: 'prompt_required'}), 'note'); input.focus();
+ throw new Error('prompt_required');
+ }
const execute = async () => {
cancelling = false;
versions.value.useOriginal = false;
@@ -186,16 +207,18 @@ export function createPromptEnhancer({node, input, editor, api, context, setText
else { versions.value.useOriginal = false; save(); render(); enhance().catch(() => {}); }
};
redo.onclick = () => enhance().catch(() => {});
- show(''); save(); render();
+ show(''); save(false); render();
return {element, async beforeGenerate() {
if (running) {
try { await running; }
catch (error) { if (versions.value.enabled && !versions.value.useOriginal) return false; }
}
- if (!versions.value.enabled || versions.value.useOriginal) return;
+ // An untouched box generates the starter prompt as it is.
+ if (!versions.value.enabled || versions.value.useOriginal || !versions.value.original.trim()) return;
try { if (!versions.fresh(context())) await enhance(); }
catch (error) { show(errorText(error), error?.message === 'cancelled' ? '' : 'error'); return false; }
}, sync() { if (input.value !== versions.text()) changed(); }, dispose() {
disposed = true; cancelJob().catch(() => {}); input.removeEventListener('input', changed);
+ input.removeEventListener('beforeinput', replace);
}};
}
diff --git a/web/result_actions.css b/web/result_actions.css
index 935bd9e..05fda82 100644
--- a/web/result_actions.css
+++ b/web/result_actions.css
@@ -19,5 +19,15 @@
.fv-actions .fv-menu-item strong{font-size:13px;font-weight:600}
.fv-actions .fv-menu-item small{font-size:11px;color:var(--fv-muted,#a5b3c6);white-space:nowrap}
.fv-actions .fv-menu-item:disabled{opacity:.7;cursor:progress}
+.fv-actions .fv-action:disabled{opacity:.65;cursor:default;transform:none}
+.fv-actions .fv-action-delete>summary:hover,.fv-actions .fv-action-delete[open]>summary{color:var(--fv-danger,#f09aa2)}
+.fv-actions .fv-confirm{width:288px;padding:14px;display:grid;gap:6px;text-align:left;cursor:default}
+.fv-actions .fv-confirm strong{font-size:13px;font-weight:600;color:var(--fv-ink,#eee)}
+.fv-actions .fv-confirm small{font-size:12px;line-height:1.55;color:var(--fv-muted,#a5b3c6);white-space:normal}
+.fv-actions .fv-confirm .fv-confirm-problem{color:var(--fv-danger,#f09aa2)}
+.fv-actions .fv-confirm-actions{display:flex;justify-content:flex-end;gap:8px;margin-top:8px}
+.fv-actions .fv-confirm-actions .fv-action{height:34px;min-height:34px;padding:0 14px}
+.fv-actions .fv-action-danger{border-color:transparent;background:var(--fv-danger-subtle,#30212a);color:var(--fv-danger,#f09aa2)}
+.fv-actions .fv-action-danger:hover{border-color:color-mix(in srgb,var(--fv-danger,#f09aa2) 45%,transparent);background:var(--fv-danger-subtle,#30212a);color:var(--fv-danger,#f09aa2)}
@keyframes fv-menu-enter{from{opacity:0;transform:translateY(4px)}to{opacity:1;transform:none}}
@media(prefers-reduced-motion:reduce){.fv-actions *{transition:none!important;animation:none!important}}
diff --git a/web/result_actions.js b/web/result_actions.js
index 17ce058..ca9cb68 100644
--- a/web/result_actions.js
+++ b/web/result_actions.js
@@ -10,6 +10,8 @@ const ICONS = {
share: '<path d="M8 10V2.5m0 0L5.2 5.3M8 2.5l2.8 2.8M4.5 7.5H4a1 1 0 0 0-1 1v4a1 1 0 0 0 1 1h8a1 1 0 0 0 1-1v-4a1 1 0 0 0-1-1h-.5"/>',
report: '<path d="M9.5 2H5a1 1 0 0 0-1 1v10a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1V4.5L9.5 2Z"/><path d="M9.5 2v2.5H12M6.5 8h3M6.5 10.5h3"/>',
chevron: '<path d="m4.5 6.5 3.5 3 3.5-3"/>',
+ expand: '<path d="M9.5 2.5h4v4M13.5 2.5 9 7M6.5 13.5h-4v-4M2.5 13.5 7 9"/>',
+ trash: '<path d="M3 4.5h10M6.5 4.5V3h3v1.5M4.5 4.5l.6 8.1a1 1 0 0 0 1 .9h3.8a1 1 0 0 0 1-.9l.6-8.1M7 7v4M9 7v4"/>',
};
export function icon(name) {
@@ -18,20 +20,93 @@ export function icon(name) {
return span;
}
+// Marks a first-pass preview on its picture, with the preview's own size.
+export function previewBadge(record, t) {
+ const badge = document.createElement('span'); badge.className = 'fv-preview-badge';
+ const g = record.preview_geometry;
+ badge.textContent = g?.width ? t(`First-pass preview · ${g.width} × ${g.height}`, `一采预览 · ${g.width} × ${g.height}`)
+ : t('First-pass preview', '一采预览');
+ return badge;
+}
+
function action(tag, label, name, cls) {
const e = document.createElement(tag); e.className = 'fv-action ' + cls;
const text = document.createElement('span'); text.textContent = label;
- e.append(icon(name), text);
+ if (name) e.append(icon(name));
+ e.append(text);
if (tag === 'button') e.type = 'button';
return e;
}
+// A popover above its button that closes on an outside click or Escape, like a
+// native menu. Escape must not also reach the dialog, which would close the window.
+function popover(menu, summary, onClose = () => {}) {
+ const outside = event => { if (menu.open && !menu.contains(event.target)) menu.open = false; };
+ menu.addEventListener('toggle', () => {
+ if (menu.open) document.addEventListener('pointerdown', outside, true);
+ else { document.removeEventListener('pointerdown', outside, true); onClose(); }
+ });
+ menu.addEventListener('keydown', event => {
+ if (event.key !== 'Escape' || !menu.open) return;
+ event.preventDefault(); event.stopPropagation(); menu.open = false; summary.focus();
+ });
+}
+
+// The next step of a first-pass preview. `run` reports its own errors and throws.
+function secondPassAction(t, run) {
+ const button = action('button', t('Run second pass', '继续二采'), 'expand', 'fv-action-primary fv-action-second-pass');
+ const label = button.lastElementChild;
+ button.title = t('Continue from this preview: latent upscale, the second pass and a full-resolution decode.',
+ '从这个预览接着跑:潜空间放大、二采,再以全分辨率解码。');
+ button.onclick = async () => {
+ if (button.disabled) return;
+ button.disabled = true;
+ try { await run(); label.textContent = t('Queued', '已加入队列'); }
+ catch { button.disabled = false; }
+ };
+ return button;
+}
+
+// Delete asks in a popover first. `remove` throws an Error whose message is shown there.
+function deleteAction(t, remove) {
+ const menu = document.createElement('details'); menu.className = 'fv-action-menu fv-action-delete';
+ const summary = document.createElement('summary'); summary.className = 'fv-action fv-action-secondary';
+ const label = document.createElement('span'); label.textContent = t('Delete', '删除');
+ summary.append(icon('trash'), label);
+ const panel = document.createElement('div'); panel.className = 'fv-menu fv-confirm';
+ const title = document.createElement('strong'); title.textContent = t('Delete this creation?', '删除这条作品?');
+ const hint = document.createElement('small');
+ hint.textContent = t('The video, its generation records and preview data are removed. This cannot be undone.',
+ '视频、生成记录和一采中间结果都会删除,无法恢复。');
+ const problem = document.createElement('small'); problem.className = 'fv-confirm-problem'; problem.hidden = true;
+ const row = document.createElement('div'); row.className = 'fv-confirm-actions';
+ const cancel = action('button', t('Cancel', '取消'), null, 'fv-action-secondary');
+ const confirm = action('button', t('Delete', '删除'), 'trash', 'fv-action-danger');
+ cancel.onclick = () => { menu.open = false; summary.focus(); };
+ confirm.onclick = async () => {
+ confirm.disabled = cancel.disabled = true; confirm.lastElementChild.textContent = t('Deleting…', '正在删除…');
+ try { await remove(); menu.open = false; }
+ catch (error) { problem.textContent = error.message; problem.hidden = false; }
+ finally { confirm.disabled = cancel.disabled = false; confirm.lastElementChild.textContent = t('Delete', '删除'); }
+ };
+ row.append(cancel, confirm);
+ panel.append(title, hint, problem, row);
+ menu.append(summary, panel);
+ popover(menu, summary, () => { problem.hidden = true; });
+ menu.addEventListener('toggle', () => { if (menu.open) cancel.focus(); });
+ return menu;
+}
+
// Download the video first; share and the reports follow, one entry each.
-// `diagnostic` is an async callback for the redacted support report, when one exists.
-export function resultActions(record, t, {diagnostic = null} = {}) {
+// A first-pass preview leads with `secondPass` instead, and saved creations end
+// with `remove`. `diagnostic` is an async callback for the redacted support
+// report, when one exists.
+export function resultActions(record, t, {diagnostic = null, secondPass = null, remove = null} = {}) {
const bar = document.createElement('div'); bar.className = 'fv-actions';
+ const finish = () => { if (remove) bar.append(deleteAction(t, remove)); return bar; };
+ if (secondPass) bar.append(secondPassAction(t, secondPass));
if (record.video) {
- const download = action('a', t('Download video', '下载视频'), 'download', 'fv-action-primary');
+ const download = action('a', t('Download video', '下载视频'), 'download', secondPass ? 'fv-action-secondary' : 'fv-action-primary');
download.href = outputDownloadURL(api, record.video); download.download = '';
bar.append(download);
}
@@ -47,7 +122,7 @@ export function resultActions(record, t, {diagnostic = null} = {}) {
report.href = recordLink; report.download = record.report.split('/').pop(); report.title = t('Generation record (JSON)', '生成记录(JSON)');
bar.append(report);
}
- return bar;
+ return finish();
}
const menu = document.createElement('details'); menu.className = 'fv-action-menu';
const summary = document.createElement('summary'); summary.className = 'fv-action fv-action-secondary';
@@ -79,20 +154,7 @@ export function resultActions(record, t, {diagnostic = null} = {}) {
list.append(full);
}
menu.append(summary, list);
- // Close on an outside click or Escape, like a native menu. Escape must not
- // also reach the dialog, which would close the whole window.
- const outside = event => { if (menu.open && !menu.contains(event.target)) menu.open = false; };
- menu.addEventListener('toggle', () => {
- if (menu.open) document.addEventListener('pointerdown', outside, true);
- else {
- document.removeEventListener('pointerdown', outside, true);
- if (!support.disabled) supportTitle.textContent = supportLabel;
- }
- });
- menu.addEventListener('keydown', event => {
- if (event.key !== 'Escape' || !menu.open) return;
- event.preventDefault(); event.stopPropagation(); menu.open = false; summary.focus();
- });
+ popover(menu, summary, () => { if (!support.disabled) supportTitle.textContent = supportLabel; });
bar.append(menu);
- return bar;
+ return finish();
}
diff --git a/web/studio.css b/web/studio.css
index 859e017..2c486b2 100644
--- a/web/studio.css
+++ b/web/studio.css
@@ -49,11 +49,24 @@
.fv-studio .fv-two-pass:has(input:disabled){opacity:.5;cursor:default}.fv-studio .fv-two-pass input:disabled{cursor:default}
.fv-studio .fv-reuse-result{display:flex;align-items:center;justify-content:center;gap:12px;color:var(--fv-muted);font-size:13px}
.fv-studio .fv-reuse-result[hidden]{display:none}
+.fv-studio .fv-preview-badge{position:absolute;top:12px;left:12px;z-index:2;padding:3px 10px;border-radius:999px;font-size:12px;line-height:1.5;font-weight:var(--fv-medium);color:var(--fv-accent);background:#0b121bcc;border:1px solid color-mix(in srgb,var(--fv-accent) 55%,transparent);backdrop-filter:blur(6px);pointer-events:none}
+.fv-studio .fv-sampling-settings .fv-pass-row{cursor:default;gap:10px}
+.fv-studio .fv-pass-row .fv-two-pass-status{margin-left:auto;font-size:11px;color:var(--fv-muted)}
+.fv-studio .fv-pass-mode{display:inline-flex;flex:none;gap:2px;padding:2px;border:1px solid var(--fv-border);border-radius:calc(var(--fv-r-sm) + 2px);background:var(--fv-bg)}
+.fv-studio .fv-pass-mode button{min-height:24px;height:24px;padding:0 10px;border:1px solid transparent;border-radius:var(--fv-r-sm);background:transparent;color:var(--fv-muted);font-size:12px;font-weight:var(--fv-medium);white-space:nowrap;box-shadow:none;cursor:pointer;transition:background .2s,color .2s}
+.fv-studio .fv-pass-mode button[aria-checked=true]{background:var(--fv-raised);color:var(--fv-ink);font-weight:var(--fv-semibold);box-shadow:0 1px 3px #0006}
+.fv-studio .fv-pass-mode button:not(:disabled):hover{color:var(--fv-ink)}
+.fv-studio .fv-pass-mode button:disabled{opacity:.4;cursor:default}
+.fv-studio .fv-pass-mode button[hidden]{display:none}
+.fv-studio .fv-pass-mode button:focus-visible{outline:2px solid var(--fv-accent);outline-offset:1px}
+.fv-studio .fv-preview-result{color:var(--fv-ink)}
@media(prefers-reduced-motion:reduce){.fv-studio .fv-two-pass,.fv-studio .fv-two-pass input,.fv-studio .fv-two-pass input::before{transition:none}}
.fv-studio .fv-preview-column{min-width:0;min-height:0;display:flex;flex-direction:column;padding:26px 30px 24px;gap:19px;overflow:auto;scrollbar-width:thin;background:radial-gradient(ellipse at 50% 30%,#6db8fa12,transparent 70%)}
.fv-studio .fv-preview-head{display:flex;align-items:center;justify-content:space-between;font-size:var(--fv-micro);color:var(--fv-muted)}
.fv-studio .fv-preview-head strong{font-size:var(--fv-body);font-weight:var(--fv-medium);color:var(--fv-ink)}
-.fv-studio .fv-preview-space{min-height:260px;flex:1 0 260px;display:grid;place-items:center;min-width:0}
+.fv-studio .fv-preview-space{min-height:260px;flex:1 0 260px;display:grid;place-items:center;min-width:0;max-height:max(260px,calc(100cqw / var(--fv-ratio,1.7778)));margin-top:auto}
+.fv-studio .fv-preview-column{container-type:inline-size}
+.fv-studio .fv-preview-end{margin-top:auto;flex:none}
.fv-studio .fv-stage{display:grid;place-items:center;position:relative;isolation:isolate;overflow:hidden;border:1px solid #ffffff14;border-radius:var(--fv-r-md);background:#101d2d;box-shadow:0 18px 52px #0004;max-width:100%;transition:width .48s cubic-bezier(.22,.7,.2,1),height .48s cubic-bezier(.22,.7,.2,1)}
.fv-studio .fv-preview-scene{position:absolute;inset:0;width:100%;height:100%;pointer-events:none;transition:opacity .55s ease}
.fv-studio .fv-stage-media{position:absolute;inset:0;display:grid;place-items:center;transition:opacity .55s ease;z-index:1}
diff --git a/web/studio.js b/web/studio.js
index 0969752..f4a512f 100644
--- a/web/studio.js
+++ b/web/studio.js
@@ -12,8 +12,9 @@ import { openLibrary, latestVideo } from './library.js';
import { createStudioQueue, randomSeed } from './studio_queue.js';
import { attachReferencePicker, referenceItems, syncReferencePrompt } from './prompt_references.js';
import { createSamplingEffort } from './sampling_effort.js';
-import { resultActions } from './result_actions.js';
+import { resultActions, previewBadge } from './result_actions.js';
import { createPromptEnhancer } from './prompt_enhance.js';
+import { rememberPromptDraft } from './prompt_draft.js';
const languageOverride = typeof location !== 'undefined'
? new URLSearchParams(location.search).get('freevideo_lang') : null;
@@ -150,8 +151,15 @@ export function openStudio(node) {
const field = (name, input) => { const label = el('label', null, 'fv-field'); label.append(el('span', name), input); return label; };
const expand = label => { const d = el('details', null, 'fv-section'); d.append(el('summary', label, 'fv-section-title')); const content = el('div', null, 'fv-expand'); d.append(content); controls.append(d); cleanup.push(animateDetails(d)); return [d, content]; };
syncReferencePrompt(node, t);
- const prompt = el('textarea'); prompt.value = value(node, 'text') || ''; prompt.placeholder = t('Describe the scene… Type @ to reference media', '描述画面… 输入 @ 引用素材'); prompt.setAttribute('aria-label', t('Prompt', '提示词'));
- prompt.disabled = linked(node, 'text'); prompt.oninput = () => set(node, 'text', prompt.value);
+ // An untouched node shows its starter prompt as a placeholder and generates it.
+ // Once something was typed, an emptied box stays empty and asks for a scene.
+ const starter = node.freevideoStarterPrompt || '';
+ const untouched = () => Boolean(starter) && value(node, 'text') === starter;
+ const shown = text => starter && text === starter ? '' : String(text || '');
+ const prompt = el('textarea'); prompt.value = shown(value(node, 'text')); prompt.setAttribute('aria-label', t('Prompt', '提示词'));
+ const placeholder = () => { prompt.placeholder = untouched() ? starter : t('Describe the scene… Type @ to reference media', '描述画面… 输入 @ 引用素材'); };
+ placeholder();
+ prompt.disabled = linked(node, 'text'); prompt.oninput = () => { set(node, 'text', prompt.value); placeholder(); };
const promptEditor = el('div', null, 'fv-prompt-editor'); promptEditor.append(prompt);
const references = attachReferencePicker(prompt, {items: () => referenceItems(node), t, view, host: dialog,
addMedia: () => { mediaDetails.open = true; mediaDetails.scrollIntoView({block: 'nearest', behavior: 'smooth'}); mediaMount.querySelector('button')?.focus(); }});
@@ -159,15 +167,16 @@ export function openStudio(node) {
mention.title = t('Reference media (@)', '引用素材(@)'); mention.setAttribute('aria-label', mention.title);
mention.disabled = prompt.disabled; mention.onpointerdown = e => e.preventDefault(); promptEditor.append(mention);
cleanup.push(() => references.dispose());
- const enhancer = createPromptEnhancer({node, input: prompt, editor: promptEditor, api, t,
+ const enhancer = createPromptEnhancer({node, input: prompt, editor: promptEditor, api, t, starter,
onReport: row => { node.freevideoReportId = row.report_id; progress.updateReport(row); },
- setText: text => set(node, 'text', text), context: () => {
+ onChange: () => rememberPromptDraft(api, node),
+ setText: text => { set(node, 'text', text || (untouched() ? starter : '')); placeholder(); }, context: () => {
const items = referenceItems(node);
if (linked(node, 'conditioning') || items.some(row => !row.file || row.kind !== 'image')) throw new Error('unsupported_media');
return {seconds: Number(value(node, 'seconds')), media: items.map(row => ({file: row.file, role: row.role || 'reference'}))};
}});
cleanup.push(() => enhancer.dispose());
- const promptChanged = e => { if (String(e.detail) === String(node.id)) { if (prompt.value !== value(node, 'text')) prompt.value = value(node, 'text') || ''; references.refresh(); enhancer.sync(); } };
+ const promptChanged = e => { if (String(e.detail) === String(node.id)) { if (prompt.value !== shown(value(node, 'text'))) prompt.value = shown(value(node, 'text')); placeholder(); references.refresh(); enhancer.sync(); } };
window.addEventListener('freevideo-reference-prompt', promptChanged);
cleanup.push(() => window.removeEventListener('freevideo-reference-prompt', promptChanged));
const promptSection = section(t('Describe your scene', '描述画面'));
@@ -199,17 +208,35 @@ export function openStudio(node) {
const seconds = el('input'); seconds.type = 'number'; seconds.min = '1.625'; seconds.max = '60'; seconds.step = 'any'; seconds.value = value(node, 'seconds'); seconds.disabled = linked(node, 'seconds'); seconds.setAttribute('aria-label', t('Duration in seconds', '时长(秒)'));
const fields = el('div', null, 'fv-fields'); fields.append(field(t('Total pixels', '总像素'), mp), field(t('Duration · seconds', '时长 · 秒'), seconds)); canvas.append(fields);
const canvasNote = el('div', null, 'fv-canvas-note'), dimensions = el('span'), duration = el('span'); canvasNote.append(dimensions, duration); canvas.append(canvasNote);
+ // State of the two-pass choice below; the radio group is what is shown.
const twoPass = el('input'); twoPass.type = 'checkbox'; twoPass.checked = value(node, 'two_pass') !== false;
- twoPass.setAttribute('role', 'switch');
twoPass.disabled = linked(node, 'two_pass') || !widget(node, 'two_pass');
// Two-pass switched off only because first-pass steps left 8; returning to 8 restores it.
let autoSinglePass = false;
- twoPass.onchange = () => { autoSinglePass = false; set(node, 'two_pass', twoPass.checked); syncSamplingSteps(); };
- const twoPassLabel = el('label', null, 'fv-two-pass');
+ const twoPassLabel = el('div', null, 'fv-two-pass fv-pass-row');
twoPassLabel.title = t('Generate at a lower resolution, then upscale and finish sampling at the target size to save time.', '先以低分辨率生成,再放大到目标分辨率完成采样,缩短生成时间。');
- twoPass.title = twoPassLabel.title;
const twoPassStatus = el('small', '', 'fv-two-pass-status');
- twoPassLabel.append(el('span', t('Two-pass acceleration', '二次采样加速')), twoPassStatus, twoPass);
+ // Off, on, or on and paused after the first pass: a preview is one way to run two-pass.
+ const passMode = el('div', null, 'fv-pass-mode'); passMode.setAttribute('role', 'radiogroup');
+ passMode.setAttribute('aria-label', t('Two-pass sampling', '二次采样'));
+ const passOptions = [['off', t('Off', '关闭')], ['on', t('On', '开启')], ['preview', t('Preview first', '先出预览')]].map(([mode, label]) => {
+ const option = el('button', label); option.type = 'button'; option.dataset.mode = mode; option.setAttribute('role', 'radio');
+ option.onclick = () => choosePassMode(mode);
+ return option;
+ });
+ passOptions[2].title = t('Two-pass, paused after the half-resolution first pass so you can look at it. Run the second pass from the result when it looks right: the video matches generating it directly.',
+ '二次采样跑完半分辨率的一采先停下,给你看预览;满意后在结果里点「继续二采」,成片与直接生成一致。');
+ passMode.append(...passOptions);
+ passMode.addEventListener('keydown', event => {
+ const step = {ArrowRight: 1, ArrowDown: 1, ArrowLeft: -1, ArrowUp: -1}[event.key];
+ if (!step) return;
+ event.preventDefault();
+ const enabled = passOptions.filter(option => !option.disabled && !option.hidden);
+ const current = enabled.findIndex(option => option.getAttribute('aria-checked') === 'true');
+ const next = enabled[(current + step + enabled.length) % enabled.length];
+ if (next) { next.click(); next.focus(); }
+ });
+ twoPassLabel.append(el('span', t('Two-pass', '二次采样')), twoPassStatus, passMode);
let effortEstimate = {}, effortEstimateKey = '', effortEstimateTimer = null, effortEstimateRequest = null;
const effort = createSamplingEffort(t, {onChange: tier => {
baseSteps.value = tier.steps;
@@ -221,6 +248,24 @@ export function openStudio(node) {
syncSamplingSteps();
}, onPreview: tier => tier ? showEffortEstimate(tier.steps, tier.twoPass) : showEffortEstimate()});
const samplingSettings = el('div', null, 'fv-sampling-settings');
+ const previewFirst = el('input'); previewFirst.type = 'checkbox'; previewFirst.checked = value(node, 'preview') === true;
+ function choosePassMode(mode) {
+ const wantsTwoPass = mode !== 'off', wantsPreview = mode === 'preview';
+ if (twoPass.checked !== wantsTwoPass) { twoPass.checked = wantsTwoPass; autoSinglePass = false; set(node, 'two_pass', wantsTwoPass); }
+ if (previewFirst.checked !== wantsPreview) { previewFirst.checked = wantsPreview; set(node, 'preview', wantsPreview); }
+ syncSamplingSteps(); updateRunButton();
+ }
+ function syncPassMode() {
+ const mode = !twoPass.checked ? 'off' : previewFirst.checked ? 'preview' : 'on';
+ for (const option of passOptions) {
+ const selected = option.dataset.mode === mode;
+ option.setAttribute('aria-checked', String(selected)); option.tabIndex = selected ? 0 : -1;
+ }
+ passOptions[0].disabled = linked(node, 'two_pass') || !widget(node, 'two_pass');
+ passOptions[1].disabled = twoPass.disabled;
+ passOptions[2].disabled = twoPass.disabled || linked(node, 'preview') || !widget(node, 'preview');
+ passOptions[2].hidden = !widget(node, 'preview');
+ }
samplingSettings.append(effort.element, twoPassLabel);
canvas.append(samplingSettings); cleanup.push(() => effort.dispose());
cleanup.push(() => { clearTimeout(effortEstimateTimer); effortEstimateRequest?.abort(); });
@@ -273,6 +318,8 @@ export function openStudio(node) {
syncSamplingSteps();
baseSteps.reportValidity();
};
+ // The run button exists only further down; until then its label is set there.
+ let runButtonReady = false;
function syncSamplingSteps() {
const baseLinked = linked(node, 'base_steps');
// Two-pass is offered at 8 first-pass steps. A saved workflow that already
@@ -281,6 +328,13 @@ export function openStudio(node) {
twoPass.disabled = unavailable || linked(node, 'two_pass') || !widget(node, 'two_pass');
twoPassStatus.textContent = unavailable ? t('Coming soon', '即将推出') : '';
twoPassLabel.classList.toggle('fv-two-pass-unavailable', unavailable);
+ // A preview stops after the first pass, so it needs two-pass sampling.
+ if (!twoPass.checked && previewFirst.checked) {
+ previewFirst.checked = false; set(node, 'preview', false);
+ if (runButtonReady) updateRunButton();
+ }
+ previewFirst.disabled = !twoPass.checked || linked(node, 'preview') || !widget(node, 'preview');
+ syncPassMode();
baseSteps.disabled = baseLinked || !widget(node, 'base_steps');
refineSteps.disabled = !twoPass.checked || linked(node, 'refine_steps') || !widget(node, 'refine_steps');
baseSteps.min = twoPass.checked ? '2' : '1';
@@ -355,7 +409,7 @@ export function openStudio(node) {
const previewHead = el('div', null, 'fv-preview-head'); previewHead.append(el('strong', t('Preview', '预览')));
const previewCaption = el('span'), previewTools = el('div', null, 'fv-preview-tools');
previewTools.append(previewCaption, button(t('Creations', '作品'), () => {
- stage.querySelector('video')?.pause(); openLibrary(t);
+ stage.querySelector('video')?.pause(); openLibrary(t, {upscale: value => node.freevideoUpscaleResult(value)});
})); previewHead.append(previewTools); output.append(previewHead);
const previewSpace = el('div', null, 'fv-preview-space'); output.append(previewSpace);
const stage = el('div', null, 'fv-stage');
@@ -373,6 +427,9 @@ export function openStudio(node) {
}
cleanup.push(hideProgress);
let previewRatio = Number(initialWidth) / Number(initialHeight);
+ // The space takes no more height than the picture can use at full width.
+ const ratioStyle = () => { if (Number.isFinite(previewRatio) && previewRatio > 0) previewSpace.style.setProperty('--fv-ratio', String(previewRatio)); };
+ ratioStyle();
function fitPreview() {
const {width, height} = previewSpace.getBoundingClientRect();
if (!width || !height || !Number.isFinite(previewRatio) || previewRatio <= 0) return;
@@ -382,7 +439,7 @@ export function openStudio(node) {
}
function previewSize(w, h) {
if (!(w > 0 && h > 0)) return;
- previewRatio = w / h;
+ previewRatio = w / h; ratioStyle();
dimensionText(previewCaption, `${w} × ${h}`);
fitPreview();
}
@@ -425,8 +482,15 @@ export function openStudio(node) {
}, 'fv-quiet');
reuseRow.append(reuseNotice, regenerate);
const trimNote = el('div', '', 'fv-note fv-trim-note'); trimNote.hidden = true; trimNote.setAttribute('role', 'status');
+ // Run second pass is the preview's primary action in the result bar.
+ const secondPass = saved => async () => {
+ try { await node.freevideoUpscaleResult(saved); await syncQueue(); }
+ catch (error) { status.dataset.error = 'true'; status.textContent = error.message; throw error; }
+ };
output.append(status, reuseRow, trimNote, links, stats, budget, progress.report, prewarm);
const failure = createErrorPanel(t); output.append(failure.element);
+ // With the picture, the result's actions and numbers form one block, centred below the header.
+ output.append(el('div', null, 'fv-preview-end'));
if (node.freevideoFailure) failure.show(node.freevideoFailureReport || node.freevideoFailure, false);
let result = node.freevideoLastResult || app.nodeOutputs?.[node.id]?.freevideo_summary?.[0];
function showResult(r) {
@@ -435,7 +499,10 @@ export function openStudio(node) {
progress.updateReport({report_id: node.freevideoReportId});
stage.querySelector('video')?.pause(); stageMedia.replaceChildren();
const video = el('video'); video.src = view(r.video, 'output'); video.controls = true; video.preload = 'metadata'; video.playsInline = true; stageMedia.append(video);
- stats.hidden = !!r.result_cache_hit;
+ // A preview is marked on the picture itself; its timings describe only
+ // the half-resolution first pass, so they stay in the report.
+ if (r.preview) stageMedia.append(previewBadge(r, t));
+ stats.hidden = !!r.result_cache_hit || !!r.preview;
reuseRow.hidden = !r.result_cache_hit;
const trims = Array.isArray(r.reference_trims) ? r.reference_trims : [];
trimNote.textContent = trims.map(row => referenceTrimText(row, t)).join('\n'); trimNote.hidden = !trims.length;
@@ -444,12 +511,13 @@ export function openStudio(node) {
metricLabels[2].textContent = unified ? t('Unified memory', '统一内存总量') : t('VRAM peak', '显存峰值');
metricLabels[3].textContent = unified ? t('Process RAM peak', '进程内存峰值') : t('RAM peak', '内存峰值');
const shown = [number(r.sample_seconds), number(r.request_seconds), number(unified ? r.unified_total_bytes : r.vram_peak_bytes, 2 ** 30, 'GiB'), number(r.ram_peak_bytes, 2 ** 30, 'GiB')]; metrics.forEach((e, i) => e.textContent = shown[i]);
- budget.textContent = r.result_cache_hit ? '' : unified
+ budget.textContent = r.result_cache_hit || r.preview ? '' : unified
? (Number.isFinite(r.unified_reserve_bytes) ? `${t('Reserved unified memory', '预留统一内存')} ${number(r.unified_reserve_bytes, 2 ** 30, 'GiB')}` : '')
: (Number.isFinite(r.gpu_budget_bytes) ? `${t('VRAM budget', '可用显存预算')} ${number(r.gpu_budget_bytes, 2 ** 30, 'GiB')} · ${t('Device', '显卡总量')} ${number(r.gpu_total_bytes, 2 ** 30, 'GiB')}` : '');
// One report entry: the diagnostic export joins the report menu once a result exists.
progress.report.dataset.covered = 'true';
- links.replaceChildren(resultActions(r, t, {diagnostic: progress.hasReport() ? progress.downloadReport : null}));
+ links.replaceChildren(resultActions(r, t, {diagnostic: progress.hasReport() ? progress.downloadReport : null,
+ secondPass: r.preview ? secondPass(r) : null}));
const g = r.geometry; if (g?.width && g?.height) previewSize(g.width, g.height);
if (!progress.element.hidden) {
progress.update({phase: 'complete', overall: {status: 'complete', fraction: 1, remaining_seconds: 0}});
@@ -499,6 +567,7 @@ export function openStudio(node) {
runOptions.append(optionsToggle, optionsPanel);
const generate = button('', submitDraft, 'fv-primary');
runrow.append(generate, runOptions);
+ runButtonReady = true;
const queueActions = el('div', null, 'fv-queue-actions');
const queueDetails = el('details', null, 'fv-queue-details');
const queueToggle = el('summary'), queueList = el('div', null, 'fv-queue-list');
@@ -542,7 +611,8 @@ export function openStudio(node) {
generate.textContent = submitting ? t('Adding…', '正在添加…')
: runMode === 'batch' ? t(`Add ${count.value || '…'} videos`, `添加 ${count.value || '…'} 条任务`)
: runMode === 'loop' ? (queueState?.looping ? t('Looping', '循环中') : t('Start loop', '开始循环'))
- : busy ? t('Add to queue', '添加到队列') : t('Generate video', '生成视频');
+ : busy ? t('Add to queue', '添加到队列')
+ : previewFirst.checked && !previewFirst.disabled ? t('Generate preview', '生成预览') : t('Generate video', '生成视频');
for (const b of modeButtons) b.setAttribute('aria-pressed', String(b.dataset.mode === runMode));
countField.hidden = runMode !== 'batch'; count.disabled = runMode !== 'batch';
modeHelp.hidden = runMode === 'single';
@@ -606,7 +676,7 @@ export function openStudio(node) {
const editing = dialog.contains(document.activeElement) ? document.activeElement : null;
if (editing && editing !== generate) { editing.blur(); editing.focus({preventScroll: true}); }
if ([...dialog.querySelectorAll('input:not(:disabled)')].some(e => !e.reportValidity())) return;
- if (!prompt.disabled && !prompt.value.trim()) { prompt.focus(); status.textContent = t('Describe your scene first.', '请先描述画面。'); return; }
+ if (!prompt.disabled && !prompt.value.trim() && !untouched()) { prompt.focus(); status.textContent = t('Describe your scene first.', '请先描述画面。'); return; }
const chosenMode = runMode, batchCount = chosenMode === 'batch' ? Number(count.value) : 1;
const originalSeed = value(node, 'seed'), seedMode = mode.value;
try {
diff --git a/web/studio_queue.js b/web/studio_queue.js
index 35ae77f..8ff4e67 100644
--- a/web/studio_queue.js
+++ b/web/studio_queue.js
@@ -26,6 +26,33 @@ export async function regenerateResult(api, nodeId, result, currentPrompt) {
api.dispatchCustomEvent?.('promptQueued', {number: 0, batchCount: 1});
return reply.prompt_id;
}
+// Upscale a preview: queue the request that produced it with refine_from set.
+// The bridge continues the preview's own saved request, so later editor
+// changes cannot alter what is upscaled.
+export async function upscaleResult(api, nodeId, result, currentPrompt) {
+ nodeId = String(nodeId);
+ if (!result?.preview || !result.video) throw new Error('not_preview');
+ const path = result.request_id ? '/history/' + encodeURIComponent(result.request_id) : '/history?max_items=64';
+ let rows = [];
+ try {
+ const response = await api.fetchApi(path);
+ if (response.ok) rows = Object.values(await response.json());
+ } catch { /* Cleared/unavailable history: the bridge still uses the preview's request. */ }
+ rows.sort((a, b) => (b.prompt?.[0] || 0) - (a.prompt?.[0] || 0));
+ const saved = rows.find(row => row.outputs?.[nodeId]?.freevideo_summary?.[0]?.video === result.video
+ && row.status?.status_str === 'success' && row.status?.completed
+ && row.prompt?.[2]?.[nodeId]?.class_type === 'FreeVideoGenerate');
+ const snapshot = saved
+ ? {output: clone(saved.prompt[2]), workflow: clone(saved.prompt[3]?.extra_pnginfo?.workflow || {})}
+ : clone(await currentPrompt());
+ if (!snapshot.output?.[nodeId]) throw new Error('missing_node');
+ Object.assign(snapshot.output[nodeId].inputs, {preview: false, refine_from: result.video, force_regenerate: false});
+ const reply = await api.queuePrompt(0, snapshot);
+ if (!reply?.prompt_id) throw new Error('submit_failed');
+ api.dispatchCustomEvent?.('promptQueued', {number: 0, batchCount: 1});
+ return reply.prompt_id;
+}
+
export function randomSeed() {
const bytes = crypto.getRandomValues(new Uint32Array(2));
return (bytes[0] & 0x1fffff) * 0x100000000 + bytes[1];