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.

Detailed generation progress with stage words and real percentages

Bowen Xueb4c46ed767ac
diff --git a/freevideo_engine/comfy_bridge.py b/freevideo_engine/comfy_bridge.py
index dcf2fcf..59eb37c 100644
--- a/freevideo_engine/comfy_bridge.py
+++ b/freevideo_engine/comfy_bridge.py
@@ -429,7 +429,7 @@ class WholeVideoProgress:
if self.request['flow'] == 'upscale' or self.retry_reuse_first:
return 'pass2'
if self.request['flow'] != 'preview' and self.request['two_pass']:
- if done > base or (done == base and (name == 'sampling_start' or message.get('stage') in ('layers', 'inputs'))):
+ if done > base or (done == base and (name == 'sampling_start' or message.get('stage') in ('layers', 'inputs', 'kernel_prepare'))):
return 'pass2'
return 'pass1'
return {'encoding': 'encode', 'load': 'load', 'decode': 'decode_save'}.get(phase) if isinstance(phase, str) else None
@@ -493,9 +493,10 @@ class WholeVideoProgress:
self.live_calibration = True
elif phase == 'load':
if event.get('event') == 'loaded':
- done, total = 50, 50
+ total = self.count[1] if self.count else 50
+ done = total
elif event.get('event') == 'prepared_blocks':
- done, total = event.get('blocks'), 50
+ done, total = message.get('done'), message.get('total')
elif event.get('event') == 'model_load_phase':
done, total = event.get('done'), event.get('total')
# LoRA groups describe a substage, not completion of the whole load.
@@ -1031,9 +1032,14 @@ def progress_message(event):
return {'label': 'Preparing reference media', 'timing_phase': 'encoding',
'reference_trimmed': {key: event.get(key) for key in ('kind', 'number', 'seconds', 'used_seconds')}}
if name == 'prepared_blocks':
- # The cached video model has 50 blocks, as in the former label.
- return {'label': 'Loading the video model',
- 'timing_phase': 'load', 'done': event.get('blocks'), 'total': 50}
+ # Old engine logs omitted the total for the pinned 50-block model.
+ # Current workers report the actual model length, including the last
+ # block when that length is not a multiple of five.
+ result = {'label': 'Loading the video model', 'timing_phase': 'load'}
+ done, total = event.get('blocks'), event.get('total', 50)
+ if type(done) is int and type(total) is int and total > 0 and 0 <= done <= total:
+ result.update(done=done, total=total)
+ return result
if name in ('sampling_preset_download', 'reference_assets_download'):
reference = name == 'reference_assets_download'
return dict(label='Preparing reference media resources' if reference else 'Preparing sampling preset',
diff --git a/freevideo_engine/decode.py b/freevideo_engine/decode.py
index 1827f3c..cedb222 100644
--- a/freevideo_engine/decode.py
+++ b/freevideo_engine/decode.py
@@ -10,21 +10,32 @@ from .offload import LayerOffloader, pin_layer_weights
from .vae_tiles import TileDecoder, compile_blocks
+DECODE_LABELS = {
+ 'vae_load': 'Loading video VAE',
+ 'video_decode': 'Decoding the video',
+ 'video_postprocess': 'Processing video colors',
+ 'rgb_save': 'Saving decoded video',
+ 'audio_load_decode': 'Loading and decoding audio',
+ 'audio_save': 'Saving decoded audio',
+ 'mp4_save': 'Saving MP4 and audio',
+}
+
+
@torch.no_grad()
def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, prefetch=True, tile_group=False,
preload=False, pin_weights=False, artifacts_dir=None, phase_callback=None,
stream_output=False, stream_weights=False, resident_blocks=0, model_cache=None,
linear_compute_cache=False, continuation=None, audio_lead_in_callback=None, output_size=None):
+ def phase(name):
+ if phase_callback is not None:
+ phase_callback(name)
+ print(json.dumps({'event': 'decode_phase', 'phase': DECODE_LABELS[name]}), flush=True)
+ phase('vae_load')
from diffusers import AutoencoderKLMiniMaxH3
from diffusers.utils.export_utils import encode_video
from src.inference.render import PIXEL_MEAN, PIXEL_STD, FPS
started = time.perf_counter()
- def phase(name):
- if phase_callback is not None:
- phase_callback(name)
- phase('vae_load')
- print(json.dumps({'event': 'decode_phase', 'phase': 'Loading video VAE'}), flush=True)
vae = None
weight_source = None
compute_cache_metrics = None
@@ -117,7 +128,6 @@ def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, pre
load_seconds = time.perf_counter() - started
decode_started = time.perf_counter()
phase('video_decode')
- print(json.dumps({'event': 'decode_phase', 'phase': 'Decoding the video'}), flush=True)
offload_stats = None
rgb_mapping = None
streamed_timings = {}
@@ -150,10 +160,10 @@ def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, pre
gc.collect()
torch.cuda.empty_cache()
video_decode_seconds = streamed_timings.get('video_decode_seconds', time.perf_counter() - decode_started)
- phase('video_postprocess')
# Bound postprocessing memory by frame chunks; use the same FP32 operations
# and rounding before converting to uint8. Do this on CPU after decoding.
if not stream_output:
+ phase('video_postprocess')
frames = torch.empty((video.shape[2], video.shape[3], video.shape[4], 3), dtype=torch.uint8)
pixel_mean = torch.tensor(PIXEL_MEAN).view(1, -1, 1, 1, 1)
pixel_std = torch.tensor(PIXEL_STD).view(1, -1, 1, 1, 1)
@@ -179,7 +189,6 @@ def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, pre
artifact_seconds += time.perf_counter() - tick
audio_start = time.perf_counter()
phase('audio_load_decode')
- print(json.dumps({'event': 'decode_phase', 'phase': 'Loading and decoding audio'}), flush=True)
audio_vae = None
if model_cache is not None:
from .resident_models import key, file_identity
@@ -195,6 +204,7 @@ def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, pre
audio_vae.eval().requires_grad_(False)
audio_mean = torch.tensor(audio_vae.config.latents_mean, device='cuda').view(1, -1, 1)
audio_std = torch.tensor(audio_vae.config.latents_std, device='cuda').view(1, -1, 1)
+ print(json.dumps({'event': 'decode_phase', 'phase': 'Decoding audio'}), flush=True)
audio = audio_vae.decode(audio_latents * audio_std + audio_mean, return_dict=False)[0]
audio = audio.float().permute(1, 0, 2)[0].cpu()
rate = audio_vae.config.sampling_rate
@@ -225,7 +235,6 @@ def decode_to_file(latents, audio_latents, out_path, *, base, offload=False, pre
temporary = destination.with_name(destination.stem + '.partial' + destination.suffix)
encode_started = time.perf_counter()
phase('mp4_save')
- print(json.dumps({'event': 'decode_phase', 'phase': 'Saving MP4 and audio'}), flush=True)
if output_size is not None:
# The model rendered a larger size; only the picture is scaled.
from .output_fit import fit_frames
diff --git a/freevideo_engine/model_load_progress.py b/freevideo_engine/model_load_progress.py
new file mode 100644
index 0000000..9b7e0d0
--- /dev/null
+++ b/freevideo_engine/model_load_progress.py
@@ -0,0 +1,20 @@
+"""Real completed video-model blocks; no tensor reads or device imports."""
+import json
+
+
+class PreparedBlocks:
+ def __init__(self, total):
+ if type(total) is not int or total < 1:
+ raise ValueError('Model block total must be a positive integer')
+ self.total, self.done = total, 0
+
+ def report(self, **fields):
+ print(json.dumps(dict(fields, event='prepared_blocks', blocks=self.done, total=self.total)), flush=True)
+
+ def complete(self, **fields):
+ # Called only after a block's existing preparation and upload submission.
+ # A submission count does not claim synchronized GPU upload completion.
+ if self.done == self.total:
+ raise ValueError('All model blocks have already been prepared')
+ self.done += 1
+ self.report(**fields)
diff --git a/freevideo_engine/progress_timeline.py b/freevideo_engine/progress_timeline.py
index 7fb3870..5b49243 100644
--- a/freevideo_engine/progress_timeline.py
+++ b/freevideo_engine/progress_timeline.py
@@ -84,6 +84,7 @@ STAGES = frozenset((
'gpu',
'imports',
'inputs',
+ 'kernel_prepare',
'installation',
'latent_save',
'latent_upscale',
@@ -123,6 +124,7 @@ LABEL_KEYS = frozenset((
'Continuing from the preview',
'Copy local model',
'Decoding video and audio',
+ 'Decoding audio',
'Decoding video tiles',
'Downloading extra model weights (one-time download)',
'Encoding image/video references',
@@ -183,6 +185,8 @@ LABEL_KEYS = frozenset((
'Preparing two-pass upscaler',
'Preparing video',
'Preparing video decoding',
+ 'Preparing GPU programs for the first step',
+ 'Processing video colors',
'Preview saved',
'Prompt ready',
'RAM measurement',
@@ -214,6 +218,8 @@ LABEL_KEYS = frozenset((
'Saving MP# and audio',
'Saving completed sampling',
'Saving prompt cache',
+ 'Saving decoded video',
+ 'Saving decoded audio',
'Single-pass · # steps',
'Start ComfyUI',
'Starting ComfyUI',
diff --git a/freevideo_engine/runtime.py b/freevideo_engine/runtime.py
index 172dae1..64188b4 100644
--- a/freevideo_engine/runtime.py
+++ b/freevideo_engine/runtime.py
@@ -292,6 +292,9 @@ class Engine:
block_bind_seconds = 0.
upload_submit_seconds = 0.
block_order = list(range(len(model.transformer_blocks)))
+ from .model_load_progress import PreparedBlocks
+ block_progress = PreparedBlocks(len(block_order))
+ block_progress.report()
if streamed_source is not None and streamed_source.direct_read:
# Pins and views are granted in the order blocks are prepared, and
# the blocks RAM cannot keep are read from the drive every step.
@@ -360,15 +363,9 @@ class Engine:
# unpinned bytes are reopened by SafetensorLayers on demand,
# rather than accumulating until the first sampling step.
prepare_streamed(block, streamed_refiner_count + index - resident_blocks, unread)
- # A block whose sampling table is computed here, for a step count
- # without stored tables, takes far longer than a cached one; report
- # each of those so the load visibly moves.
- if prepared % 5 == 0 or (adaln_cache and not cache_hit):
- print(json.dumps({'event': 'prepared_blocks', 'blocks': prepared,
- 'adaln_cache': adaln_cache, 'adaln_reused_blocks': table_cache_hits,
- 'resident_uploaded_blocks': min(prepared, resident_blocks),
- 'phase': 'Load cached blocks and upload resident weights',
- 'seconds': time.perf_counter() - started}), flush=True)
+ block_progress.complete(adaln_cache=adaln_cache, adaln_reused_blocks=table_cache_hits,
+ resident_uploaded_blocks=min(prepared, resident_blocks),
+ phase='Load cached blocks and upload resident weights', seconds=time.perf_counter() - started)
if adaln_cache:
del embeddings
self.schedule_hook = model.register_forward_pre_hook(self.cursor.before, with_kwargs=True)
diff --git a/freevideo_engine/sampling_progress.py b/freevideo_engine/sampling_progress.py
index 63558cd..17b1169 100644
--- a/freevideo_engine/sampling_progress.py
+++ b/freevideo_engine/sampling_progress.py
@@ -89,6 +89,9 @@ class SamplingProgress:
def start(self):
self.event('sampling_start', stage='inputs', step=1)
+ # Preparation can encounter a new shape or a cold compiler cache. We
+ # cannot tell whether a compile is active, or how far it has progressed.
+ self.event('sampling_prepare', stage='kernel_prepare', step=1)
def layer(self, index):
if len(self.completed) >= self.steps:
@@ -160,6 +163,15 @@ def progress_message(event):
return dict(label=labels[phase], phase='sample_finalize', stage=phase,
done=0, total=0)
return None
+ if name == 'sampling_prepare':
+ done, total = event.get('completed_steps'), event.get('total')
+ if type(total) is not int or total < 1 or type(done) is not int or not 0 <= done < total:
+ return None
+ return dict(label='Preparing GPU programs for the first step', phase='sampling',
+ stage='kernel_prepare', done=done, total=total, unit='steps',
+ kernel_cache_note=True,
+ uniform_remaining_steps=event.get('uniform_remaining_steps', True),
+ step_elapsed_seconds=event.get('step_elapsed_seconds', 0.))
if name not in ('sampling_start', 'sampling_progress', 'step', 'loaded'):
return None
total = event.get('total', 8)
diff --git a/scripts/test_freevideo_entry.mjs b/scripts/test_freevideo_entry.mjs
index 53fe143..bc4f4ea 100644
--- a/scripts/test_freevideo_entry.mjs
+++ b/scripts/test_freevideo_entry.mjs
@@ -351,11 +351,12 @@ test('main browser entry registers both views and preserves node hooks', async t
const activity = progress.snapshot().label;
removalNode.freevideoShowProgress({compatibility: {level: 2, name_en: 'Compatible', name_zh: '兼容'}});
assert.equal(progress.snapshot().label, activity, 'The node must not replace the activity for settings-only events');
- assert.match(progress.snapshot().status, /Compatibility setting Compatible|兼容档/);
+ assert.match(progress.snapshot().status, /so far|已进行/, 'pause has priority over a settings-only compatibility notice');
removalNode.freevideoShowProgress({phase: 'recovery', retry: {kind: 'ram_pressure'}});
assert.match(progress.snapshot().status, /Out of RAM|内存不足/);
removalNode.freevideoShowProgress({new_request: true, phase: 'load', received_at: 0});
- assert.equal(progress.snapshot().status, '', 'A new request clears the old retry and compatibility notice');
+ assert.match(progress.snapshot().status, /so far|已进行/, 'An aged new request clears retry and compatibility but shows its own pause');
+ assert.doesNotMatch(progress.snapshot().status, /Out of RAM|内存不足|Compatibility setting|兼容档/);
assert.equal(progress.stall.hidden, false);
assert.equal(timers.size, 1);
let layoutCalls = 0, timersDuringLayout;
diff --git a/web/editor.css b/web/editor.css
index b537f53..15e8494 100644
--- a/web/editor.css
+++ b/web/editor.css
@@ -63,7 +63,11 @@
.fv-ed-genbar{grid-column:1/-1;grid-row:4;display:flex;align-items:center;gap:12px;padding:0 16px;border-top:1px solid var(--fv-border);border-bottom:1px solid var(--fv-border);background:var(--fv-card);min-width:0}
.fv-ed-genbar .fv-ed-description{width:auto}
.fv-ed-clip[data-state=running] .fv-block-progress.fv-ed-block-progress{display:grid}
-.fv-ed-gen-hint{font-size:13px;color:var(--fv-muted);flex:1;min-width:0}
+.fv-ed-gen-hint{display:flex;font-size:13px;color:var(--fv-muted);flex:1;min-width:0;overflow:hidden;white-space:pre}
+.fv-ed-gen-part{flex:none}
+.fv-ed-gen-stage,.fv-ed-gen-time{min-width:0;overflow:hidden;text-overflow:ellipsis}
+.fv-ed-gen-stage{flex:0 1000 auto}
+.fv-ed-gen-time{flex:none;max-width:100%}
.fv-ed-genbar[data-mode=selected] .fv-ed-gen-hint{flex:none}
.fv-ed-gen-note{font-size:13px;color:var(--fv-ink);flex:0 1 auto;min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
.fv-ed-gen-note[hidden]{display:none}
diff --git a/web/editor.js b/web/editor.js
index 2efdc99..7f3211b 100644
--- a/web/editor.js
+++ b/web/editor.js
@@ -209,7 +209,8 @@ export function openEditor({api, app, from = null, panel = {}, components: given
function barJob(job) {
const latest = [...runProgress.values()].filter(message => message.chain.id === job.id)
.sort((a, b) => (a.received_at || 0) - (b.received_at || 0)).at(-1);
- return latest ? {...job, chainProgress: latest.chain} : job;
+ const words = panel.progressWords?.() || null;
+ return latest ? {...job, chainProgress: latest.chain, words} : words ? {...job, words} : job;
}
function paintJob(job) {
const messages = blockMessages([...runProgress.values()].filter(message => message.chain.id === job.id),
@@ -1254,7 +1255,10 @@ export function openEditor({api, app, from = null, panel = {}, components: given
listeners.push(() => api.removeEventListener('execution_error', executionError));
}
if (panel.progressElement && typeof MutationObserver !== 'undefined') {
- progressObserver = new MutationObserver(() => { for (const job of jobStatuses.values()) paintJob(job); });
+ progressObserver = new MutationObserver(() => {
+ for (const job of jobStatuses.values()) paintJob(job);
+ if (mostRecentJob && jobStatuses.has(mostRecentJob)) genbar?.update(barJob(jobStatuses.get(mostRecentJob)));
+ });
progressObserver.observe(panel.progressElement, {subtree: true, childList: true, characterData: true, attributes: true, attributeFilter: ['aria-valuenow', 'hidden']});
}
const bar = tracks.element.querySelector('.fv-ed-track-zoom');
diff --git a/web/editor_entry.js b/web/editor_entry.js
index 3f2feb8..2420faa 100644
--- a/web/editor_entry.js
+++ b/web/editor_entry.js
@@ -29,6 +29,17 @@ export function openEditorView(node, {studio, work = null, extend = false} = {})
const number = progressElement.querySelector('.fv-generation-count')?.textContent?.match(/(\d+)%/);
return number ? {percent: Number(number[1])} : {percent: 0};
};
+ // What the creation panel is saying now; the editor covers that panel, so its bar repeats the words.
+ panel.progressWords = () => {
+ if (progressElement.hidden) return null;
+ const words = selector => {
+ const element = progressElement.querySelector(selector);
+ return element && !element.hidden && element.getClientRects().length ? element.textContent.replace(/\s+/g, ' ').trim() : '';
+ };
+ // The block already shows the overall percentage; the bar keeps only the stage sentence.
+ const stage = words('.fv-generation-label').replace(/(本步 \d+%)$| \(\d+% of this step\)$/, '');
+ return {stage, time: words('.fv-generation-status'), stall: words('.fv-generation-stall')};
+ };
}
const handle = openEditor({api, app, panel, from: work ? {work, ...(extend ? {extend: true} : {})} : null, onView(view) {
if (view === 'nodes') showNodes(node, studio);
diff --git a/web/editor_genbar.js b/web/editor_genbar.js
index 009c856..137a2cf 100644
--- a/web/editor_genbar.js
+++ b/web/editor_genbar.js
@@ -90,8 +90,17 @@ export function createEditorGenbar(host, {state, t, actions, text = ''}) {
const blocks = status?.blocks || [], done = blocks.filter(block => block.state === 'done').length;
const number = Math.min(blocks.length || 1, Math.max(done, status?.chainProgress?.index || 0) + 1), count = blocks.length || 1;
const remaining = status?.chainProgress?.overall?.remaining_seconds ?? (blocks.every(block => Number.isFinite(block.remaining_seconds)) ? blocks.reduce((sum, block) => sum + block.remaining_seconds, 0) : null);
- hint.textContent = t(`Part ${number}/${count}`, `第 ${number}/${count} 段`)
- + (remaining != null ? ' · ' + timeEstimate(remaining, t) : '');
+ const words = status?.words;
+ // The creation panel's own sentences, so a long step says what it is and how long it has taken.
+ // At a narrow width only the stage sentence is cut; the part number and the time stay.
+ // A single part needs no number; with nothing else to say it reads "Preparing".
+ const time = words?.stall || words?.time || (remaining != null ? timeEstimate(remaining, t) : '');
+ const stage = words?.stage || (count === 1 && !time ? t('Preparing', '正在准备') : '');
+ const spans = [['fv-ed-gen-part', count > 1 ? t(`Part ${number}/${count}`, `第 ${number}/${count} 段`) : ''],
+ ['fv-ed-gen-stage', stage], ['fv-ed-gen-time', time]];
+ const shown = spans.filter(([, text]) => text).map(([cls, text], index) => [cls, index ? ` · ${text}` : text]);
+ hint.replaceChildren(...shown.map(([cls, text]) => el('span', cls, text)));
+ hint.title = shown.map(([, text]) => text).join('');
}
firstFrame.update(busy);
references?.update(busy, busy ? undefined : !!actions.references.visible());
diff --git a/web/generation_progress.js b/web/generation_progress.js
index 9c3b5d9..e6314f0 100644
--- a/web/generation_progress.js
+++ b/web/generation_progress.js
@@ -11,6 +11,64 @@ const duration = value => {
};
const progressFields = ['phase', 'stage', 'timing_phase', 'label', 'done', 'total', 'block'];
+// Iris's approved substage sentences. Unknown engine labels never reach the UI.
+const SUBACTIVITY = {
+ start: ['Starting the generation process', '正在启动生成进程'],
+ libraries: ["Loading the text encoder's libraries", '正在载入文本编码器的程序库'],
+ weights: ['Reading the text encoder from disk', '正在从硬盘读取文本编码器'],
+ gpu: ['Moving the text encoder to the GPU', '正在把文本编码器载入显卡'],
+ prompt: ['Encoding the prompt', '正在编码提示词'],
+ reference: ['Encoding the reference images and videos', '正在编码参考图片和视频'],
+ load: ['Loading the video model', '正在载入视频模型'],
+ first: ['Running the first step (it takes longer the first time a size is used or after an update)',
+ '正在进行第一步(首次使用这个尺寸或更新后,第一步要多花一些时间)'],
+ upscale: ['Upscaling the first pass', '正在放大第一遍的结果'],
+ decoder: ['Loading the video decoder', '正在载入视频解码器'],
+ decode: ['Decoding the video', '正在解码视频'],
+ audio: ['Decoding the audio', '正在解码音频'],
+ save: ['Saving the video', '正在保存视频'],
+};
+// [EN, ZH, EN with a step, ZH with a step]; a request that samples twice adds which pass.
+const SAMPLING_NAME = {
+ single: ['Sampling', '正在采样', 'Sampling step', '正在采样第 '],
+ first: ['Sampling (first pass)', '正在采样(一采)', 'Sampling step', '正在采样第 '],
+ pass2: ['Sampling (second pass)', '正在采样(二采)', 'Sampling step', '正在采样第 '],
+ preview: ['Sampling (preview)', '正在采样(预览)', 'Sampling step', '正在采样第 '],
+};
+const SAMPLING_PASS = {first: ['first pass', '一采'], pass2: ['second pass', '二采'], preview: ['preview', '预览']};
+const SUBACTIVITY_LABELS = {
+ 'Starting generation process': 'start', 'Starting text encoder': 'libraries',
+ 'Loading text encoder': 'libraries', 'Preparing text encoder GPU': 'gpu',
+ 'Reading text encoder weights': 'weights', 'Preparing text encoder model': 'weights',
+ 'Loading text encoder onto GPU': 'gpu', 'Encoding prompt': 'prompt', 'Encoding text and images': 'prompt',
+ 'Loading video VAE': 'decoder', 'Loading and decoding audio': 'audio', 'Saving MP4 and audio': 'save',
+ 'Decoding the video': 'decode', 'Processing video colors': 'decode', 'Decoding audio': 'audio',
+ 'Saving decoded video': 'save', 'Saving decoded audio': 'save',
+};
+function subactivity(message, activity, t) {
+ const phase = message.phase || message.timing_phase || message.stage;
+ const label = message.label;
+ let key = SUBACTIVITY_LABELS[label];
+ const sampling = ['pass1', 'pass2', 'preview'].includes(activity?.key);
+ if (sampling) {
+ // Engines send kernel_prepare before every pass's first step; its count is the first pass's steps.
+ if (message.done === 0 || message.stage === 'kernel_prepare') key = 'first';
+ // A paused stage line carries no counts; the normal line keeps the user's steps.
+ else return {key: `sampling:${activity.key}`, text: t(...SAMPLING_NAME.single.slice(0, 2)), sampling: true};
+ }
+ key ||= {prepare: phase === 'starting' ? 'start' : null, encode: LABEL_ACTIVITY.reference.includes(label)
+ ? 'reference' : 'prompt', load: 'load', decode: 'decode', save: 'save', upscale: 'upscale'}[activity?.key];
+ if (key) return {key, text: t(...SUBACTIVITY[key]), sampling};
+ // Only a download name needs the fallback frame; every other activity already reads as a sentence.
+ const named = activity && activity.key !== 'prepare' ? activity.text : '';
+ return {key: `other:${activity?.key || 'prepare'}`, sampling,
+ text: !named ? t(...ACTIVITY.prepare) : activity.key === 'download' ? t(`Still working: ${named}`, `仍在进行:${named}`) : named};
+}
+const waitDuration = (seconds, t) => {
+ const n = Math.floor(Math.max(0, seconds)), minutes = Math.floor(n / 60), rest = n % 60;
+ return minutes ? t(`${minutes} min${rest ? ` ${rest} s` : ''}`, rest ? `${minutes} 分 ${rest} 秒` : `${minutes} 分钟`)
+ : t(`${rest} s`, `${rest} 秒`);
+};
// Only our status sentences, scoped to a request; no prompt or engine log.
// A reopened Studio/tab must not announce the replayed explanation again.
const announcementKey = 'freevideo.progress.announcements';
@@ -37,18 +95,17 @@ const STALL_SECONDS_BY_PHASE = Object.freeze({
latent_upscale: 180, sample_finalize: 180, decode: 180, save: 180,
dependencies: 180, recovery: 180, other: 180,
});
-// Only these fixed stage sentences reach either progress view, with at most the
-// sampling step or a download's size and speed. Engine labels, layer, block and
-// frame counts and the original sampling plan remain available in the report.
+// Normal stage sentences keep the sampling step or a download's size and speed.
+// After 15 s of engine silence SUBACTIVITY names the work without counts.
+// Raw engine labels and internal units remain available in the report.
// Stage words the product already uses: settings say 一采 / 二采, results say 解码.
-// No "正在": the stage line sits beside a moving clock and a small bar.
const ACTIVITY = {
- prepare: ['Preparing', '准备中'],
- prompt: ['Encoding the prompt', '编码提示词'],
- combined: ['Encoding the prompt and references', '编码提示词和参考素材'],
- load: ['Loading the video model', '载入视频模型'],
- decode: ['Decoding the video', '解码视频'],
- save: ['Saving the video', '保存视频'],
+ prepare: ['Preparing', '正在准备'],
+ prompt: ['Encoding the prompt', '正在编码提示词'],
+ combined: ['Encoding the prompt and references', '正在编码提示词和参考素材'],
+ load: ['Loading the video model', '正在载入视频模型'],
+ decode: ['Decoding the video', '正在解码视频'],
+ save: ['Saving the video', '正在保存视频'],
};
const LABEL_ACTIVITY = {
prepare: ['Starting generation process', 'Preparing video', 'Checking saved video', 'Resource forecast ready'],
@@ -68,9 +125,10 @@ const LABEL_ACTIVITY = {
'Low-memory mode for this GPU'],
refine: ['Upscaling before the second pass', 'Continuing from the preview', 'Reusing the completed first pass'],
decode: ['Loading video VAE', 'Decoding video and audio', 'Decoding the video', 'Decoding video tiles',
- 'Loading and decoding audio', 'Reusing completed sampling · retrying video and audio decoding',
+ 'Loading and decoding audio', 'Processing video colors', 'Decoding audio',
+ 'Reusing completed sampling · retrying video and audio decoding',
'Checking completed sampling', 'Saving completed sampling', 'Releasing sampling buffers', 'Preparing video decoding'],
- save: ['Saving MP4 and audio'],
+ save: ['Saving MP4 and audio', 'Saving decoded video', 'Saving decoded audio'],
};
// Engine steps that run only when the request has reference media.
const REFERENCE_WORK = new Set(['Encoding reference media', 'Encoding image/video references', 'Encoding reference audio',
@@ -103,13 +161,18 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = f
}
return {key: 'download', text: t(download[0], download[1]) + detail};
}
- const steps = (key, en, zh, step, total) => ({key, text: Number.isInteger(step) && Number.isInteger(total) && total > 0
- ? t(`${en} · step ${step} of ${total}`, `${zh} · 第 ${step} / ${total} 步`) : t(en, zh)});
+ const steps = (key, name, step, total) => {
+ const [en, zh, enSteps, zhSteps] = SAMPLING_NAME[name];
+ const pass = SAMPLING_PASS[name];
+ return {key, text: Number.isInteger(step) && Number.isInteger(total) && total > 0
+ ? t(`${enSteps} ${step} of ${total}${pass ? ` (${pass[0]})` : ''}`, `${zhSteps}${step} / ${total} 步${pass ? `(${pass[1]})` : ''}`)
+ : t(en, zh)};
+ };
const base = Number.isInteger(plan.base_steps) && plan.base_steps > 0 ? plan.base_steps : null;
const second = plan.second || plan.upscale_target;
if (message.prediction || message.sampling_plan) return named('prepare');
if (phase === 'recovery') return message.retry?.reuse_sampling === true ? named('decode')
- : message.retry?.reuse_first_pass === true ? steps('pass2', 'Second pass', '二采') : named('prepare');
+ : message.retry?.reuse_first_pass === true ? steps('pass2', 'pass2') : named('prepare');
if (phase === 'sample_finalize') return named('decode');
if (message.stage === 'latent_upscale' || phase === 'latent_upscale')
return {key: 'upscale', text: Number.isInteger(second?.width) && Number.isInteger(second?.height)
@@ -123,12 +186,12 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = f
if ((secondPass || phase === 'refinement') && twoPass) {
const refine = total !== null && total > base ? total - base : plan.refine_steps;
const step = done === null ? null : Math.min(Math.max(1, done - base + 1), refine);
- return steps('pass2', 'Second pass', '二采', step, refine);
+ return steps('pass2', 'pass2', step, refine);
}
const count = twoPass ? base : total;
const step = done === null || count === null ? null : Math.min(done + 1, count);
- if (twoPass && plan.preview) return steps('preview', 'Preview', '预览', step, count);
- return twoPass ? steps('pass1', 'First pass', '一采', step, count) : steps('pass1', 'Sampling', '采样', step, count);
+ if (twoPass && plan.preview) return steps('preview', 'preview', step, count);
+ return steps('pass1', twoPass ? 'first' : 'single', step, count);
}
const key = labelActivities.get(message.label) || {
starting: 'prepare', prepare: 'prepare', prediction: 'prepare', encoding: 'combined',
@@ -136,8 +199,11 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = f
}[phase] || labelActivities.get(message.stage);
if (!key) return null;
if (key === 'refine') return {key: 'upscale', text: t('Second pass · upscaling', '二采 · 放大')};
- // One sentence for the whole encoding stage. The engine names images in its
- // steps either way; only a request with reference media says so.
+ // Named encoder and post-sampling events stay distinct, even before a pause.
+ const substage = SUBACTIVITY_LABELS[message.label] || (key === 'reference' ? 'reference' : null);
+ if (substage) return {key: ['prompt', 'combined', 'reference'].includes(key) ? 'encode' : key,
+ text: t(...SUBACTIVITY[substage])};
+ // Older or unnamed encoding events keep the safe generic activity.
if (key === 'prompt' || key === 'combined' || key === 'reference')
return {key: 'encode', text: t(...ACTIVITY[references ? 'combined' : 'prompt'])};
// LoRA preparation counts LoRA files, then the model counts its own blocks:
@@ -322,6 +388,8 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
let pauseAnnounced = false, lastExplanation = '', announcedMessages = new Set();
let announcementRecord = null;
let remainingSeconds = null;
+ let detail = null, localShare = null, decodeMinutes = null, decodeMinutesAt = -Infinity;
+ const workShares = new Map();
let completedSteps = 0, previousRemaining = null, remainingAt = 0, lastActivity = null;
// Once the shown time reaches zero before the video is saved, it stays hidden for this request.
// The last minute runs on real time even while an estimate holds the number.
@@ -390,7 +458,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
if (id === reportId) report.textContent = t('Report unavailable · retry', '报告暂不可用 · 重试');
} finally { if (id === reportId) report.disabled = false; }
};
- const counted = value => Number.isInteger(value.total) && value.total > 0 && Number.isInteger(value.done)
+ const hasCount = value => Number.isInteger(value.total) && value.total > 0 && Number.isInteger(value.done)
&& value.done >= 0 && value.done <= value.total;
function drawOverall() {
@@ -436,25 +504,23 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
// A segment can take several minutes. Extended silence gets a notice once;
// leave the run going, and
// offer the report, whose engine log shows the step it stopped at.
- let stallHasReport = null, stallLoading = null;
+ let stallHasReport = null;
+ const segmentTypical = () => detail?.sampling && valid(state.estimated_step_seconds)
+ ? state.estimated_step_seconds : overall.typical_seconds;
function drawStall(age, ended) {
const floor = STALL_SECONDS_BY_PHASE[state.phase || state.timing_phase] ?? STALL_SECONDS_BY_PHASE.other;
- const threshold = Math.max(floor, valid(overall.typical_seconds) ? 2 * overall.typical_seconds : 0);
+ const typical = segmentTypical();
+ const threshold = Math.max(floor, valid(typical) ? 2 * typical : 0);
const shown = !ended && age >= threshold;
const wasShown = !stall.hidden;
const hasReport = reportId !== null;
- // Shown after 3 minutes, without saying so: the card shows two times only.
- const loading = (state.phase || state.timing_phase) === 'load';
- if (shown && (stall.hidden || hasReport !== stallHasReport || loading !== stallLoading)) {
- const first = loading
- ? t('No loading progress for a while. Generation is still running, so you can keep waiting.',
- '已经有一段时间没有新的加载进度,生成仍在进行,可以继续等待。')
- : t('No progress for a while. Generation is still running, so you can keep waiting.',
+ // The separate notice never adds another clock or blocks the run.
+ if (shown && (stall.hidden || hasReport !== stallHasReport)) {
+ const first = t('No progress for a while. Generation is still running, so you can keep waiting.',
'已经有一段时间没有新的进度,生成仍在进行,可以继续等待。');
const sentence = hasReport ? first + t(
' You can also use Download report below and submit the report in a GitHub issue; we will look into it and work on a fix for you.',
'也可以点击下方的“下载报告”并提交到 GitHub issue,我们会专门排查和优化,为您解决这个问题。') : first;
- stallLoading = loading;
if (hasReport) {
const buttonName = el('span', ''); buttonName.style.whiteSpace = 'nowrap';
buttonName.textContent = t('Download report', '“下载报告”');
@@ -491,10 +557,12 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
let holdUntil = 0, holdMode = null, holdFraction = null;
function miniTarget() {
const key = lastActivity?.key;
- const counted = value => Number.isFinite(value?.[0]) && Number.isFinite(value?.[1]) && value[1] > 0
+ const ratio = value => Number.isFinite(value?.[0]) && Number.isFinite(value?.[1]) && value[1] > 0
? Math.max(0, Math.min(1, value[0] / value[1])) : null;
if (key === 'download' || key === 'load' || key === 'lora' || key === 'decode' || key === 'encode') {
- const share = counted([state.done, state.total]);
+ if (localShare !== null && hasCount(state) && ['load', 'decode'].includes(detail?.key))
+ return {mode: 'determinate', fraction: localShare};
+ const share = ratio([state.done, state.total]);
if (share !== null) return {mode: 'determinate', fraction: share};
// Slow without a count (an older engine, an audio pass): sweep only.
return {mode: 'indeterminate'};
@@ -517,7 +585,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
// The current stage has lasted 2 s: the stage line names it from now on.
function settle(at) {
if (stageId !== null && stageId !== labelId && at - stageSince >= 2000) {
- labelId = stageId; labelText = lastActivity?.text ?? labelText; away = null;
+ labelId = stageId; labelText = lastActivity ? withWorkPercent(lastActivity.text) : labelText; away = null;
}
}
function enterStage(id, at) {
@@ -562,17 +630,22 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
if (!lastActivity) { labelId = null; return t(...ACTIVITY.prepare); }
settle(now());
if (labelId === null) { labelId = stageId; away = null; }
- if (labelId === stageId) labelText = lastActivity.text;
+ if (labelId === stageId) labelText = withWorkPercent(lastActivity.text);
return labelText;
}
+ function withWorkPercent(base) {
+ const percent = localShare === null ? null : Math.min(localShare === 1 ? 100 : 99, Math.floor(localShare * 100 + 1e-9));
+ // Keep the percentage with its own sentence while a new stage waits to be named.
+ return percent !== null && (detail?.key === 'load' || detail?.key === 'decode')
+ ? base + t(` (${percent}% of this step)`, `(本步 ${percent}%)`) : base;
+ }
function drawActivity(age, ended) {
- const base = ended ? (overall.status === 'complete' || state.phase === 'complete' ? savedText() : t('Stopped', '已停止'))
- : stageText();
const paused = !ended && age > 15;
- // Two times only (elapsed, remaining): a pause shows on the small bar, not as text.
- if (label.textContent !== base) label.textContent = base;
+ const base = ended ? (overall.status === 'complete' || state.phase === 'complete' ? savedText() : t('Stopped', '已停止'))
+ : paused ? detail?.text || stageText() : stageText();
+ const shown = !ended && paused && detail ? withWorkPercent(base) : base;
+ if (label.textContent !== shown) label.textContent = shown;
drawMini(age, ended);
- const firstSlow = state.phase === 'sampling' && state.stage !== 'latent_upscale' && state.done === 0 && paused;
const automatic = compatibility?.origin === 'automatic' && compatibility.level > 0;
const required = Number.isInteger(compatibility?.recovery_successes_required) && compatibility.recovery_successes_required > 0
? compatibility.recovery_successes_required : 3;
@@ -583,17 +656,33 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
: t(remaining === 1 ? 'Generation is slowed automatically for stability; it speeds up after one more successful generation.'
: `Generation is slowed automatically for stability; it speeds up after ${remaining} more successful generations.`,
`为了稳定,生成已自动放慢;再连续完成 ${remaining} 次后会加快一些。`) : '';
- if (compatibilityNote.textContent !== compatibilityText) {
- compatibilityNote.textContent = compatibilityText;
- compatibilityNote.hidden = !compatibilityText;
- layout();
+ // All explanations share one row, including automatic compatibility.
+ compatibilityNote.hidden = true;
+ let pauseText = '';
+ if (paused) {
+ const elapsed = waitDuration((now() - segmentStartedAt) / 1000, t);
+ const typical = segmentTypical(), usual = valid(typical) && typical > 0 ? waitDuration(typical, t) : '';
+ const similar = overall.typical_source === 'similar';
+ pauseText = t(`${elapsed} so far`, `已进行 ${elapsed}`);
+ if (usual) pauseText += similar ? t(` (about ${usual} on similar hardware)`, `(相近配置约 ${usual})`)
+ : t(` (usually about ${usual})`, `(通常约 ${usual})`);
}
- const level = ended ? '' : recovery ? 'retry' : firstSlow ? 'first-step'
- : lowMemory ? 'memory' : memoryTight ? 'unified-memory' : compatibility && !automatic && compatibility.level > 0 ? 'compatibility' : trims.size ? 'notice' : '';
+ // Only an available engine estimate can supply this long-decode fallback.
+ const remainingEstimate = rawRemaining();
+ if (!ended && detail?.key === 'decode' && valid(remainingEstimate) && remainingEstimate >= 180) {
+ const minutes = Math.ceil(remainingEstimate / 60);
+ if (decodeMinutes === null || now() - decodeMinutesAt >= 60000 && Math.abs(minutes - decodeMinutes) >= 1) {
+ decodeMinutes = minutes; decodeMinutesAt = now();
+ }
+ } else { decodeMinutes = null; decodeMinutesAt = -Infinity; }
+ const decodeEta = decodeMinutes === null ? '' : t(`About ${decodeMinutes} min left`, `还需约 ${decodeMinutes} 分钟`);
+ const progressText = [paused ? pauseText : '', decodeEta].filter(Boolean).join(t(' · ', ','));
+ const level = ended ? '' : recovery ? 'retry' : paused ? 'pause'
+ : lowMemory ? 'memory' : memoryTight ? 'unified-memory' : automatic ? 'automatic'
+ : compatibility && compatibility.level > 0 ? 'compatibility' : trims.size ? 'notice' : '';
const explanation = {retry: recovery,
- 'first-step': t('The first run at this size, or after an update, takes longer to start.',
- '首次使用这个尺寸或更新后,开头要多花一些时间。'),
- memory: t('This generation uses low-memory mode; it is slower than usual but will finish.',
+ pause: progressText,
+ memory: t('This video uses a memory-saving mode, so it takes longer than usual but will finish.',
'这次生成使用省显存方式,会比平时慢,但能完成。'),
// Mac, level 3 with low-memory mode (never both). The measured numbers go to the
// report only. Closing applications during the first pass can still give the
@@ -605,21 +694,28 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
'可用的统一内存较少,这次生成会比平时慢。'),
compatibility: compatibility ? t(`Compatibility setting ${compatibility.name_en} is on, so generation is slower than usual.`,
`已开启兼容性设置(${compatibility.name_zh}档),生成会比平时慢。`) : '',
- notice: trimStatus([...trims.values()], t)}[level] || '';
+ automatic: compatibilityText,
+ notice: trimStatus([...trims.values()], t)}[level] || (!ended ? progressText : '');
if (status.textContent !== explanation) status.textContent = explanation;
status.hidden = !explanation; status.dataset.level = level;
const announcements = [];
- if (paused && !pauseAnnounced) { announcements.push(base); pauseAnnounced = true; }
- for (const text of [compatibilityText, explanation !== lastExplanation ? explanation : '']) {
+ const pauseKey = `pause:${segmentKey}`;
+ if (paused && !pauseAnnounced) {
+ if (!announcedMessages.has(pauseKey)) announcements.push(base);
+ announcedMessages.add(pauseKey); pauseAnnounced = true;
+ }
+ // Seconds and local percentages are visual only. Announce the work once.
+ const spoken = level === 'pause' || explanation === progressText ? '' : explanation;
+ for (const text of [spoken !== lastExplanation ? spoken : '']) {
if (!text || announcedMessages.has(text)) continue;
announcements.push(text); announcedMessages.add(text);
- if (announcementRecord) {
- announcementRecord.messages = [...announcedMessages];
- try { sessionStorage.setItem(announcementKey, JSON.stringify(Object.fromEntries(requestAnnouncements))); }
- catch { /* In-memory deduplication still works. */ }
- }
}
- lastExplanation = explanation;
+ if (announcementRecord && announcements.length) {
+ announcementRecord.messages = [...announcedMessages];
+ try { sessionStorage.setItem(announcementKey, JSON.stringify(Object.fromEntries(requestAnnouncements))); }
+ catch { /* Optional storage; in-memory deduplication remains. */ }
+ }
+ lastExplanation = spoken;
if (announcements.length) live.textContent = announcements.join(t(' ', ''));
}
@@ -735,6 +831,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
recovery = ''; lowMemory = false; compatibility = null; references = false; trims.clear();
memoryTight = null; firstPassOver = false;
samplingPlan = {}; secondPass = false; hasHeartbeat = false;
+ detail = null; localShare = null; workShares.clear(); decodeMinutes = null; decodeMinutesAt = -Infinity;
announcementRecord = null;
}
if (!placeholderReset && (message.stream_id !== undefined || terminal)) placeholder = false;
@@ -753,6 +850,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
const heartbeat = beat && previousProgress !== null;
if (message.reset === true) {
hasHeartbeat = false; lastActivity = null;
+ workShares.clear(); localShare = null; decodeMinutes = null; decodeMinutesAt = -Infinity;
// A stage restarting its own count (the sampler's start) leaves the
// bridge's whole-request estimate standing; a retry starts afresh.
if (estimateSource && !message.retry && message.phase !== 'recovery') { completedSteps = 0; previousRemaining = null; }
@@ -762,6 +860,14 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
// sampler's own start, keeps the stage's clock and the 2 s rule.
if (miniFraction > 0) resetMini();
}
+ if (!heartbeat && message.retry && message.phase === 'recovery') {
+ // Resource retries do not carry reset=true. Clear only the new attempt's work.
+ const retryId = JSON.stringify(message.retry);
+ if (retryId !== JSON.stringify(state.retry)) {
+ workShares.clear(); localShare = null; resetMini();
+ decodeMinutes = null; decodeMinutesAt = -Infinity;
+ }
+ }
if (beat) hasHeartbeat = true;
if (!heartbeat && message.phase === 'sampling' && valid(message.done))
completedSteps = Math.max(completedSteps, message.done);
@@ -826,13 +932,24 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
enterStage(activity.key === 'download' ? `download:${state.stage || state.phase}` : activity.key, receivedAt);
lastActivity = activity;
}
- const key = JSON.stringify([state.phase || state.timing_phase, state.stage,
- state.phase === 'sampling' ? [state.done, state.total] : [state.label, lastActivity?.text]]);
+ detail = subactivity(state, lastActivity, t);
+ localShare = null;
+ if (['load', 'decode'].includes(detail.key) && lastActivity?.key !== 'lora') {
+ const shareKey = `${detail.key}:${secondPass ? 'second' : 'first'}`;
+ const share = hasCount(state) ? state.done / state.total
+ : state.label === 'Video model ready' ? 1 : null;
+ if (share !== null) workShares.set(shareKey, Math.max(workShares.get(shareKey) ?? 0, share));
+ localShare = workShares.get(shareKey) ?? null;
+ }
+ const key = JSON.stringify([detail.key, detail.sampling ? [lastActivity?.key, state.done] : null]);
if (key !== segmentKey || message.reset) {
segmentKey = key;
- segmentStartedAt = receivedAt - (state.phase === 'sampling' && valid(message.step_elapsed_seconds)
+ segmentStartedAt = receivedAt - (detail.sampling && valid(message.step_elapsed_seconds)
? message.step_elapsed_seconds * 1000 : 0);
pauseAnnounced = false;
+ if (detail.sampling && !('estimated_step_seconds' in message)) delete state.estimated_step_seconds;
+ } else if (detail.sampling && valid(message.step_elapsed_seconds)) {
+ segmentStartedAt = Math.min(segmentStartedAt, receivedAt - message.step_elapsed_seconds * 1000);
}
element.dataset.phase = message.phase || message.timing_phase || '';
if (message.overall && typeof message.overall === 'object') {
@@ -850,6 +967,7 @@ export function createGenerationProgress(t, now = () => Date.now(), {compact = f
if (timer !== null) clearInterval(timer); timer = null; state = {}; overall = {}; shownFraction = null;
stopCount(); countShown = null;
resetRemaining(); lastActivity = null; resetMini(); previousProgress = null; segmentKey = ''; status.textContent = ''; live.textContent = '';
+ detail = null; localShare = null; workShares.clear(); decodeMinutes = null; decodeMinutesAt = -Infinity;
// Last: the layout callback sizes the node from the cleared state.
drawStall(0, true);
}