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.

Add advanced sampling step controls (#14)

Bowen Xuec0b1971cfff0
diff --git a/freevideo_engine/adaptive.py b/freevideo_engine/adaptive.py
index fd58052..e1c55f8 100644
--- a/freevideo_engine/adaptive.py
+++ b/freevideo_engine/adaptive.py
@@ -483,7 +483,7 @@ 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 (type(steps) is not int or steps != 8 or len(times) != total_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.')
rgb_shape, dtype = _retained_array_header(Path(artifacts)/'rgb.npy')
diff --git a/freevideo_engine/cli.py b/freevideo_engine/cli.py
index b5e3247..0890178 100644
--- a/freevideo_engine/cli.py
+++ b/freevideo_engine/cli.py
@@ -98,7 +98,11 @@ def main():
generate.add_argument('--no-tuning', action='store_true', help='Bypass saved tuning and conditioning reuse for this request')
generate.add_argument('--seed', type=int, default=2026090901)
generate.add_argument('--two-pass', action=argparse.BooleanOptionalAction, default=True,
- help='Automatic smaller-canvas 8 steps + latent upscale + 2 tail steps (default: on, including media conditioning); --no-two-pass restores full-resolution 8 steps')
+ help='Sample a smaller canvas, upscale latents and refine (default: on; 8 + 2 steps)')
+ generate.add_argument('--base-steps', type=int, choices=range(1, 33), metavar='1..32',
+ 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=2, metavar='1..31',
+ help='Second-pass tail steps; must be fewer than first-pass steps (default: 2). Changes may reduce quality.')
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_bridge.py b/freevideo_engine/comfy_bridge.py
index cd68047..3bddc2b 100644
--- a/freevideo_engine/comfy_bridge.py
+++ b/freevideo_engine/comfy_bridge.py
@@ -331,15 +331,17 @@ def progress_message(event):
if name == 'sampling_plan':
first, second = event.get('first', {}), event.get('second')
if event.get('enabled') and second:
- return dict(label='Two-pass · %d × %d → %d × %d · 8 + 2 steps' %
- (first['width'], first['height'], second['width'], second['height']),
+ return dict(label='Two-pass · %d × %d → %d × %d · %d + %d steps' %
+ (first['width'], first['height'], second['width'], second['height'],
+ event.get('base_steps', 8), event.get('refine_steps', 2)),
detail='Audio is preserved from the first pass', sampling_plan=event)
- return dict(label='Single-pass · 8 steps', detail=event.get('reason'), sampling_plan=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 == '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',
- done=8, total=10, display_fraction=.8, timing_phase='sampling',
+ 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 == 'ram_budget_warning':
@@ -415,9 +417,12 @@ def engine_environment(root, source, environ=None):
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):
+ release_models=None, export_inputs=None, two_pass=True, encoder_prewarm=None,
+ base_steps=8, refine_steps=2):
if type(two_pass) is not bool:
raise ValueError('Two-pass generation must be a boolean')
+ from .two_pass import validate_steps
+ validate_steps(base_steps, refine_steps, two_pass)
canvas = validate_request(prompt, width, height, seconds, seed)
source = Path(source or source_root()).resolve()
root, machine = installation(source, environ)
@@ -439,7 +444,8 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
command = [machine['python'], '-B', '-m', 'freevideo_engine.managed', '--root', str(root),
'generate', '--prompt-file', str(run / 'prompt.txt'), '--out', str(output),
'--width', str(width), '--height', str(height), '--seconds', str(seconds), '--seed', str(seed),
- '--two-pass' if two_pass else '--no-two-pass']
+ '--two-pass' if two_pass else '--no-two-pass',
+ '--base-steps', str(base_steps), '--refine-steps', str(refine_steps)]
if resources['gpu_reserve_gib'] is not None:
command += ['--gpu-reserve-gib', str(resources['gpu_reserve_gib'])]
state = {'status': 'starting', 'geometry': canvas, 'seed': seed, 'installation': str(root),
@@ -482,7 +488,7 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
state['command'] = command
from .media_request import task_for
from .two_pass import plan
- planned = plan(canvas, two_pass, task_for(extra.get('media', {})))
+ planned = plan(canvas, two_pass, task_for(extra.get('media', {})), base_steps=base_steps, refine_steps=refine_steps)
whole_progress.sampling_plan = planned
whole_progress.forecast(progress_history_forecast(root, machine, dict(canvas, sampling_plan=planned)))
send_progress({'label': 'Preparing %.3f s video + audio · %d × %d' %
@@ -516,7 +522,8 @@ def generate(prompt, width, height, seconds, seed, output_directory, *,
engine = json.loads(output.with_suffix('.engine.json').read_text(encoding='utf-8'))
from .two_pass import steps, plan
completed_steps = steps(report.get('sampling_plan'))
- expected_plan = plan(canvas, two_pass, report.get('profile', {}).get('engine', {}).get('task', 't2va'))
+ expected_plan = plan(canvas, two_pass, report.get('profile', {}).get('engine', {}).get('task', 't2va'),
+ base_steps=base_steps, refine_steps=refine_steps)
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
diff --git a/freevideo_engine/comfy_nodes.py b/freevideo_engine/comfy_nodes.py
index d8fe196..05b533f 100644
--- a/freevideo_engine/comfy_nodes.py
+++ b/freevideo_engine/comfy_nodes.py
@@ -41,6 +41,10 @@ class FreeVideoGenerate(io.ComfyNode):
Media.Input('media', optional=True, tooltip='Unified Media panel: keyframes or ordered references.'),
io.Boolean.Input('two_pass', display_name='Two-pass sampling', default=True, optional=True,
tooltip='Generate the scene, then refine it at the target resolution.'),
+ io.Int.Input('base_steps', display_name='First-pass steps', default=8, min=1, max=32, optional=True,
+ tooltip='Default: 8. Changing sampling steps may reduce generation quality.'),
+ io.Int.Input('refine_steps', display_name='Second-pass steps', default=2, min=1, max=31, optional=True,
+ tooltip='Default: 2. Must be fewer than first-pass steps. Changing sampling steps may reduce generation quality.'),
],
outputs=[io.Video.Output('video'), io.String.Output('report', display_name='Report JSON')],
hidden=[io.Hidden.unique_id, io.Hidden.extra_pnginfo],
@@ -51,6 +55,8 @@ 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', 2), kwargs.get('two_pass', True))
comfy_bridge.installation()
except (OSError, ValueError, KeyError) as error:
return str(error)
@@ -64,7 +70,9 @@ 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):
+ references=None, loras=None, conditioning=None, media=None, two_pass=True, base_steps=8, refine_steps=2):
+ from .two_pass import validate_steps
+ validate_steps(base_steps, refine_steps, two_pass)
from .comfy_media import export
import folder_paths
import comfy.model_management as memory
@@ -76,7 +84,7 @@ class FreeVideoGenerate(io.ComfyNode):
from .resident_process import OWNER
gpu_prewarm = OWNER.stop_prewarm()
node_id = cls.hidden.unique_id
- bar = ProgressBar(8, node_id=node_id)
+ bar = ProgressBar(base_steps + (refine_steps if two_pass else 0), node_id=node_id)
last_message = [None]
last_count = [None]
def progress(message):
@@ -103,6 +111,7 @@ class FreeVideoGenerate(io.ComfyNode):
try:
output = comfy_bridge.generate(text, width, height, seconds, seed, output_root,
metadata=metadata.get('workflow', metadata), progress=progress, two_pass=two_pass,
+ base_steps=base_steps, refine_steps=refine_steps,
encoder_prewarm=gpu_prewarm or cpu_prewarm,
interrupted=memory.throw_exception_if_processing_interrupted, release_models=release,
export_inputs=lambda run, canvas: export(run, canvas, first=first, last=last,
diff --git a/freevideo_engine/decode_resume.py b/freevideo_engine/decode_resume.py
index 52890f2..392faae 100644
--- a/freevideo_engine/decode_resume.py
+++ b/freevideo_engine/decode_resume.py
@@ -38,7 +38,8 @@ def completed_sampling(metrics):
except (ValueError, AttributeError, TypeError):
return False
return (metrics.get('finite_latents') is True and isinstance(config, dict)
- and type(config.get('steps')) is int and config['steps'] == 8
+ and type(config.get('steps')) is int
+ and config['steps'] == (metrics.get('sampling_plan') or {}).get('base_steps', 8)
and isinstance(steps, list) and len(steps) == total
and all(type(value) in (int, float) and math.isfinite(value) and value > 0 for value in steps)
and isinstance(metrics.get('sampling_provenance'), dict) and bool(metrics['sampling_provenance']))
diff --git a/freevideo_engine/generate.py b/freevideo_engine/generate.py
index 48d989e..2534ea4 100644
--- a/freevideo_engine/generate.py
+++ b/freevideo_engine/generate.py
@@ -170,6 +170,9 @@ def automatic_profile(args, canvas, *, stage, evidence, descriptor=None, environ
row.update(capacity_trial=True, resource_error=details)
row['status'] = 'admitted'
profile = selected.legacy_profile()
+ if 'steps' in canvas:
+ profile['engine']['steps'] = canvas['steps']
+ profile['policy']['engine']['steps'] = canvas['steps']
row['budget_bytes'] = dict(gpu=profile['gpu_budget_gb']*1e9, ram=profile['inference_ram_budget_gb']*1e9)
row['reserve_bytes'] = dict(gpu=profile.get('policy', {}).get('gpu_system_reserve_bytes'),
ram=profile.get('policy', {}).get('ram_system_reserve_bytes'))
@@ -266,6 +269,14 @@ def _run(args):
from .geometry import geometry
canvas = geometry(getattr(args, 'width', 1344), getattr(args, 'height', 768),
frames=getattr(args, 'frames', None), seconds=getattr(args, 'seconds', None))
+ from .two_pass import validate_steps
+ requested_steps = getattr(args, 'base_steps', None)
+ refine_steps = getattr(args, 'refine_steps', 2)
+ two_pass = getattr(args, 'two_pass', True)
+ if requested_steps is not None:
+ validate_steps(requested_steps, refine_steps, two_pass)
+ if requested_steps != 8:
+ canvas['steps'] = requested_steps
from .media_request import read as read_media, task_for, cache_key
media = read_media(getattr(args, 'media', None))
active_loras = [row for row in media.get('loras', []) if row['strength'] != 0]
@@ -294,6 +305,13 @@ def _run(args):
inference_ram_budget_gb=profile['ram_budget_bytes'] / 1e9,
engine=profile['engine'], decoder=profile['decoder'],
allocator_config=profile['allocator_config'], policy=profile)
+ base_steps = requested_steps if requested_steps is not None else profile['engine'].get('steps', 8)
+ validate_steps(base_steps, refine_steps, two_pass)
+ profile['engine']['steps'] = base_steps
+ if 'engine' in profile.get('policy', {}):
+ profile['policy']['engine']['steps'] = base_steps
+ if base_steps != 8:
+ canvas['steps'] = base_steps
if args.base:
profile['engine']['base'] = str(args.base.resolve())
if args.checkpoint:
@@ -488,12 +506,10 @@ def _run(args):
if not condition.is_file():
raise FileNotFoundError(condition)
from .two_pass import plan as plan_sampling, ensure_checkpoint
- sampling_plan = plan_sampling(canvas, getattr(args, 'two_pass', True),
- profile['engine'].get('task', 't2va'))
+ 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 sampling_plan['enabled']:
- if profile['engine'].get('steps', 8) != 8:
- raise ValueError('Automatic two-pass generation requires the original DMD8 schedule')
+ 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)
save(destination.with_suffix('.request.json'), report)
@@ -507,7 +523,7 @@ def _run(args):
hardware, resident_credit, live = automatic_profile(args, canvas, stage='after-encoding',
evidence=planning, descriptor=descriptor, environment=env)
report['reclaimable_resident_models'] = resident_credit
- live['engine'].update({k: profile['engine'][k] for k in ('base', 'checkpoint', 'task') if k in profile['engine']})
+ live['engine'].update({k: profile['engine'][k] for k in ('base', 'checkpoint', 'task', 'steps') if k in profile['engine']})
profile = live
report['profile'] = profile
from .prediction import predict
@@ -519,7 +535,7 @@ def _run(args):
tokens = next((row['observation'].get('text_tokens', row.get('geometry', {}).get('text_tokens'))
for row in reversed(observations)
if row['observation'].get('conditioning_sha256') == condition_hash), None)
- if state and args.attention == 'auto' and tokens is not None:
+ if state and args.attention == 'auto' and tokens is not None and sampling_plan['version'] == 1:
profile, applied = tuning.apply_profile(state, profile, canvas, tokens)
report['profile'] = profile
report['tuning']['profile_id'] = applied
diff --git a/freevideo_engine/optimize.py b/freevideo_engine/optimize.py
index a367ad7..437132c 100644
--- a/freevideo_engine/optimize.py
+++ b/freevideo_engine/optimize.py
@@ -75,8 +75,9 @@ def read_baseline(path, expected):
continue
preencoded = row.get('input_mode') == 'preencoded'
validate_metrics(row, row['geometry'], preencoded=preencoded)
- if row['engine'].get('sampling_plan', {}).get('enabled'):
- continue # Optimizer compares an unchanged 8-step sampler, not a two-pass output.
+ if (row['engine'].get('sampling_plan', {}).get('enabled')
+ or row['engine'].get('config', {}).get('steps') != 8):
+ continue # Optimizer comparisons require the unchanged eight-step sampler.
if not isinstance(row['id'], str) or Path(row['id']).name != row['id'] or row['id'] in ('.', '..'):
raise ValueError('Invalid baseline case identifier')
case = directory / row['id']
diff --git a/freevideo_engine/prediction.py b/freevideo_engine/prediction.py
index 90a6d18..21858f8 100644
--- a/freevideo_engine/prediction.py
+++ b/freevideo_engine/prediction.py
@@ -82,7 +82,7 @@ def _canvas(value, options, text_tokens=None):
seconds=value.get('seconds') if value.get('frames') is None else None)
canvas['steps'] = options['engine']['steps']
canvas['task'] = options['engine']['task']
- if value.get('sampling_plan', {}).get('enabled'):
+ if value.get('sampling_plan', {}).get('enabled') or value.get('sampling_plan', {}).get('version') == 2:
canvas['sampling_plan'] = value['sampling_plan']
if type(canvas['steps']) is not int or canvas['steps'] <= 0:
raise ValueError('Forecast steps must be a positive integer')
diff --git a/freevideo_engine/resource_history.py b/freevideo_engine/resource_history.py
index 14b66ad..0de809c 100644
--- a/freevideo_engine/resource_history.py
+++ b/freevideo_engine/resource_history.py
@@ -115,7 +115,7 @@ def computation_identity(identity, config, geometry, *, purpose='generation'):
raise ValueError('Durable resource history requires GPU UUID and operating system')
computation = _canonical_config(config, purpose)
canvas = {name: geometry.get(name) for name in ('width', 'height', 'frames')}
- if geometry.get('sampling_plan', {}).get('enabled'):
+ if geometry.get('sampling_plan', {}).get('enabled') or geometry.get('sampling_plan', {}).get('version') == 2:
canvas['sampling_plan'] = geometry['sampling_plan']
for key, default in (('steps', 8), ('task', 't2va')):
configured = computation['engine'].pop(key, None)
diff --git a/freevideo_engine/two_pass.py b/freevideo_engine/two_pass.py
index 779f4bb..4901080 100644
--- a/freevideo_engine/two_pass.py
+++ b/freevideo_engine/two_pass.py
@@ -1,4 +1,4 @@
-"""Automatic 8 + 2 generation planning, without importing Torch."""
+"""Generation planning with an 8 + 2 default, without importing Torch."""
from math import gcd
from pathlib import Path
@@ -11,7 +11,18 @@ UPSCALER = dict(
role='latent_upscaler')
-def plan(canvas, enabled=True, task='t2va'):
+def validate_steps(base_steps=8, refine_steps=2, enabled=True):
+ if type(enabled) is not bool:
+ raise ValueError('Two-pass generation must be a boolean')
+ if type(base_steps) is not int or not 1 <= base_steps <= 32:
+ raise ValueError('First-pass steps must be an integer between 1 and 32')
+ if type(refine_steps) is not int or not 1 <= refine_steps <= 31:
+ raise ValueError('Second-pass steps must be an integer between 1 and 31')
+ if enabled and refine_steps >= base_steps:
+ raise ValueError('Second-pass steps must be fewer than first-pass steps')
+
+
+def plan(canvas, enabled=True, task='t2va', *, base_steps=8, refine_steps=2):
"""Align the smaller canvas, lift, center-crop, then refine at the target.
Odd multiples of 32 need up to 32 extra pixels before the latent crop.
@@ -20,12 +31,14 @@ def plan(canvas, enabled=True, task='t2va'):
"""
from .geometry import geometry
from .media_request import TASKS
+ validate_steps(base_steps, refine_steps, enabled)
if task not in TASKS:
raise ValueError('Unsupported two-pass conditioning task: ' + str(task))
target = geometry(canvas['width'], canvas['height'], frames=canvas['frames'])
shape = {key: target[key] for key in ('width', 'height', 'frames')}
- result = dict(version=1, requested=bool(enabled), enabled=False, mode='vdn8',
- base_steps=8, refine_steps=0, total_steps=8, first=shape, second=None)
+ default = base_steps == 8 and (not enabled or refine_steps == 2)
+ result = dict(version=1 if default else 2, requested=enabled, enabled=False, mode='vdn%d' % base_steps,
+ base_steps=base_steps, refine_steps=0, total_steps=base_steps, first=shape, second=None)
if not enabled:
result['reason'] = 'Two-pass generation is disabled.'
return result
@@ -45,14 +58,17 @@ def plan(canvas, enabled=True, task='t2va'):
top=(lift['height'] - target['height']) // 2,
width=target['width'], height=target['height'])
upscaling = first != lift
- result.update(enabled=True, mode='vdn8-lbh-tail2', refine_steps=2, total_steps=10,
+ result.update(enabled=True, mode='vdn%d-lbh-tail%d' % (base_steps, refine_steps),
+ refine_steps=refine_steps, total_steps=base_steps + refine_steps,
first=first, second=shape, upscale_target=lift, crop=crop,
restart_seed_offset=1,
upscaler_sha256=UPSCALER['sha256'] if upscaling else None,
audio_policy='first_pass_preserved_with_audio_clock_conditioning',
- reason='8 steps on a smaller canvas, learned latent upscale, then the original DMD8 schedule tail (2 steps).')
+ reason='%d steps on a smaller canvas, learned latent upscale, then the DMD schedule tail (%d steps).' % (base_steps, refine_steps))
+ if default:
+ result['reason'] = '8 steps on a smaller canvas, learned latent upscale, then the original DMD8 schedule tail (2 steps).'
if not upscaling:
- result['reason'] = 'Small canvas: 8 steps and 2 refinement steps at the same target size, without upscaling.'
+ result['reason'] = 'Small canvas: %d steps and %d refinement steps at the same target size, without upscaling.' % (base_steps, refine_steps)
return result
@@ -72,11 +88,15 @@ def steps(sampling_plan=None):
"""Validate the supported receipt rather than trusting a supplied NFE count."""
if sampling_plan is None:
return 8
- expected = 10 if sampling_plan.get('enabled') is True else 8
- if (sampling_plan.get('version') != 1 or sampling_plan.get('base_steps') != 8
- or sampling_plan.get('refine_steps') != expected - 8
- or sampling_plan.get('total_steps') != expected):
- raise ValueError('Invalid 8 + 2 sampling plan')
+ enabled = sampling_plan.get('enabled')
+ base, refine = sampling_plan.get('base_steps'), sampling_plan.get('refine_steps')
+ validate_steps(base, refine if enabled else 2, enabled)
+ expected = base + refine if enabled else base
+ if (type(sampling_plan.get('version')) is not int or sampling_plan['version'] not in (1, 2)
+ or (sampling_plan['version'] == 1 and (base != 8 or refine != (2 if enabled else 0)))
+ or (not enabled and (type(refine) is not int or refine != 0))
+ or type(sampling_plan.get('total_steps')) is not int or sampling_plan['total_steps'] != expected):
+ raise ValueError('Invalid sampling step plan')
return expected
@@ -89,7 +109,7 @@ def same_strategy(left, right):
except (ValueError, AttributeError, TypeError):
return False
# Until a two-resolution cost model is calibrated, use exact-plan evidence.
- return steps(a) == 8 or a == b
+ return a == b if (a or {}).get('enabled') or (b or {}).get('enabled') else True
def checkpoint_path():
@@ -176,4 +196,6 @@ def first_pass_policy(profile, canvas, sampling_plan):
lora_max_block_bytes=policy.get('lora_max_block_bytes', 0),
lora_root_bytes=policy.get('lora_root_bytes', 0),
canvas=first, allow_capacity_trial=policy.get('capacity_trial', False)).legacy_profile()
+ selected['engine']['steps'] = sampling_plan['base_steps']
+ selected['policy']['engine']['steps'] = sampling_plan['base_steps']
return dict(geometry=first, profile=selected)
diff --git a/freevideo_engine/validation.py b/freevideo_engine/validation.py
index 066d9fd..fb53c13 100644
--- a/freevideo_engine/validation.py
+++ b/freevideo_engine/validation.py
@@ -41,7 +41,8 @@ def validate_metrics(row, canvas, *, preencoded=False):
steps = engine.get('step_seconds', [])
from .two_pass import steps as planned_steps
sampling_plan = row['request'].get('sampling_plan')
- if (engine.get('sampling_plan') != sampling_plan or engine.get('config', {}).get('steps') != 8
+ if (engine.get('sampling_plan') != sampling_plan
+ or engine.get('config', {}).get('steps') != (sampling_plan or {}).get('base_steps', 8)
or len(steps) != planned_steps(sampling_plan) or not all(finite_number(v) for v in steps)):
raise ValueError('Expected all planned denoising steps with finite timings and matching sampling strategy')
timings = [('request', ('request_seconds',)),
diff --git a/freevideo_engine/worker.py b/freevideo_engine/worker.py
index 389d38f..4d7210e 100644
--- a/freevideo_engine/worker.py
+++ b/freevideo_engine/worker.py
@@ -50,7 +50,9 @@ def generate(request, resident=None):
if 'sampling_plan' in request:
from .two_pass import plan
expected = plan(request['geometry'], request['sampling_plan']['requested'],
- request['engine_options'].get('task', 't2va'))
+ request['engine_options'].get('task', 't2va'),
+ base_steps=request['engine_options'].get('steps', 8),
+ refine_steps=request['sampling_plan']['refine_steps'] if request['sampling_plan']['enabled'] else 2)
if request['sampling_plan'] != expected:
raise ValueError('Sampling plan does not match the requested canvas and task')
metrics['sampling_plan'] = expected
@@ -171,8 +173,8 @@ def generate(request, resident=None):
from .two_pass import upscale_workspace, crop_latents
from .two_pass_metrics import combine
from .latent_upscale import upscale
- if engine.steps != 8:
- raise ValueError('Two-pass generation requires the original DMD8 schedule')
+ if engine.steps != sampling_plan['base_steps']:
+ raise ValueError('Engine steps disagree with the requested sampling plan')
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)
@@ -183,11 +185,12 @@ def generate(request, resident=None):
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=10,
+ 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=8, total=10)), flush=True)
+ 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')
@@ -268,10 +271,11 @@ def generate(request, resident=None):
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=2,
+ **sampling_plan['second'], initial_latents=(latents, audio), refine_steps=sampling_plan['refine_steps'],
final_step_callback=None if decoder_cached else decode_read_ahead.start,
step_callback=memory_diagnostics.complete, sampling_complete_callback=refined_complete,
- progress_offset=8, progress_total=10, gpu_reserve_bytes=request.get('gpu_reserve_bytes', 0),
+ 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)
diff --git a/web/error_panel.js b/web/error_panel.js
index e13abcb..60db2fa 100644
--- a/web/error_panel.js
+++ b/web/error_panel.js
@@ -109,7 +109,7 @@ export function createErrorReport(detail, context = {}) {
const inputs = context.snapshot?.output?.[context.node?.id]?.inputs
|| (nodeType === 'FreeVideoGenerate' ? detail?.current_inputs : null);
const settings = {};
- for (const key of ['width', 'height', 'seconds', 'seed', 'two_pass']) {
+ for (const key of ['width', 'height', 'seconds', 'seed', 'two_pass', 'base_steps', 'refine_steps']) {
const value = Array.isArray(inputs?.[key]) && inputs[key].length === 1 ? inputs[key][0] : inputs?.[key];
if (typeof value === 'boolean' || typeof value === 'number' && Number.isFinite(value)) settings[key] = value;
}
diff --git a/web/freevideo.js b/web/freevideo.js
index b2eea90..b920861 100644
--- a/web/freevideo.js
+++ b/web/freevideo.js
@@ -321,6 +321,14 @@ app.registerExtension({
quality.label = text('Two-pass sampling', '二次采样');
quality.tooltip = text('Generate the scene, then refine it at the target resolution.', '先生成画面,再以目标分辨率精修。');
}
+ for (const [name, en, zh] of [['base_steps', 'First-pass steps', '一采步数'], ['refine_steps', 'Second-pass steps', '二采步数']]) {
+ const steps = this.widgets?.find(w => w.name === name);
+ if (steps) {
+ steps.label = text(en, zh);
+ steps.tooltip = text('Changing sampling steps may reduce generation quality. Defaults: 8 + 2. Second-pass steps must be fewer than first-pass steps.',
+ '修改采样步数可能降低生成质量。默认一采 8 步、二采 2 步;二采步数必须小于一采步数。');
+ }
+ }
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/studio.js b/web/studio.js
index a55966c..9bdd29f 100644
--- a/web/studio.js
+++ b/web/studio.js
@@ -178,7 +178,7 @@ export function openStudio(node) {
const canvasNote = el('div', null, 'fv-canvas-note'), dimensions = el('span'), duration = el('span'); canvasNote.append(dimensions, duration); canvas.append(canvasNote);
const twoPass = el('input'); twoPass.type = 'checkbox'; twoPass.checked = value(node, 'two_pass') !== false;
twoPass.disabled = linked(node, 'two_pass') || !widget(node, 'two_pass');
- twoPass.onchange = () => set(node, 'two_pass', twoPass.checked);
+ twoPass.onchange = () => { set(node, 'two_pass', twoPass.checked); syncSamplingSteps(); };
const twoPassLabel = el('label', null, 'fv-two-pass');
twoPassLabel.title = t('Generate the scene, then refine it at the target resolution.', '先生成画面,再以目标分辨率精修。');
twoPass.title = twoPassLabel.title;
@@ -205,6 +205,40 @@ export function openStudio(node) {
const width = el('input'), height = el('input');
for (const [w, name, label] of [[width, 'width', t('Width', '宽度')], [height, 'height', t('Height', '高度')]]) { w.type = 'number'; w.min = '256'; w.max = '4096'; w.step = '32'; w.value = value(node, name); w.disabled = dimensionsLinked; w.setAttribute('aria-label', label); custom.append(field(label, w)); w.onchange = () => { if (w.value && w.reportValidity()) { set(node, name, Number(w.value)); selected = 'custom'; node.properties.freevideo_aspect = selected; ratio = Number(width.value) / Number(height.value); updateCanvas(); } }; }
advanced.append(custom);
+ const samplingFields = el('div', null, 'fv-fields'); samplingFields.style.marginTop = '14px';
+ const baseSteps = el('input'), refineSteps = el('input');
+ const samplingWarning = el('p', t('Changing sampling steps may reduce generation quality. Defaults: 8 + 2 steps.', '修改采样步数可能降低生成质量。默认一采 8 步、二采 2 步。'), 'fv-muted');
+ samplingWarning.id = `fv-sampling-warning-${node.id}`;
+ samplingWarning.style.color = 'var(--fv-warning)';
+ for (const [input, name, label, fallback, maximum] of [
+ [baseSteps, 'base_steps', t('First-pass steps', '一采步数'), 8, 32],
+ [refineSteps, 'refine_steps', t('Second-pass steps', '二采步数'), 2, 31],
+ ]) {
+ input.type = 'number'; input.min = '1'; input.max = String(maximum); input.step = '1'; input.required = true;
+ input.value = value(node, name) ?? fallback;
+ input.setAttribute('aria-label', label); input.setAttribute('aria-describedby', samplingWarning.id);
+ input.oninput = () => syncSamplingSteps();
+ input.onchange = () => input.reportValidity();
+ samplingFields.append(field(label, input));
+ }
+ function syncSamplingSteps() {
+ const baseLinked = linked(node, 'base_steps');
+ 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';
+ refineSteps.max = baseLinked ? '31' : String(Math.max(1, Math.min(31, Number(baseSteps.value || 8) - 1)));
+ refineSteps.setCustomValidity(twoPass.checked && !baseLinked && !refineSteps.disabled && Number(refineSteps.value) >= Number(baseSteps.value)
+ ? t('Second-pass steps must be fewer than first-pass steps.', '二采步数必须小于一采步数。') : '');
+ refineSteps.title = twoPass.checked ? t('Must be fewer than first-pass steps.', '必须小于一采步数。')
+ : t('Enable two-pass sampling to use this setting.', '开启二次采样后生效。');
+ for (const [input, name] of [[baseSteps, 'base_steps'], [refineSteps, 'refine_steps']]) {
+ if (!input.disabled && input.checkValidity() && Number(input.value) !== value(node, name)) {
+ set(node, name, Number(input.value));
+ }
+ }
+ }
+ syncSamplingSteps();
+ advanced.append(samplingFields, samplingWarning);
const seedFields = el('div', null, 'fv-fields'); seedFields.style.marginTop = '14px';
const seed = el('input'); seed.type = 'number'; seed.min = '0'; seed.max = String(2 ** 53 - 1); seed.step = '1'; seed.value = value(node, 'seed'); seed.disabled = linked(node, 'seed'); seed.setAttribute('aria-label', t('Seed', '种子'));
seed.onchange = () => { if (seed.value && seed.reportValidity()) set(node, 'seed', Number(seed.value)); };