Spaces:

ymo2017
/

latentsync-16-zero

Running on Zero

App Files Files Community

latentsync-16-zero

Commit History

Oct 10

22:50
add missing stage2.yaml for v1.5 Running

ymo2017

Verified signature
9f64dae
22:47
switch to v1.5 256px (fewer mask artifacts)

ymo2017

Verified signature
4bf9529
21:50
fix: deploy official affine_transform.py (AlignRestore resolution)

ymo2017

Verified signature
27dc1b6
21:47
fix: add missing enable_deepcache arg

ymo2017

Verified signature
8a9b057
21:31
clean rebuild: official audio2feature.py

ymo2017

Verified signature
39b69e0
21:28
fix: restore official face_detector.py

ymo2017

Verified signature
27826b8
21:27
fix: add missing check_ffmpeg_installed to util.py

ymo2017

Verified signature
d71c5b7
21:21
fix: unpin torch/torchvision for ZeroGPU

ymo2017

Verified signature
589810b
21:20
fix: loosen huggingface-hub for ZeroGPU gradio compat

ymo2017

Verified signature
8d5d077
21:18
fix: remove gradio pin to avoid ZeroGPU conflict

ymo2017

Verified signature
b5d5700
20:48
fix: use ZeroGPU-compatible torch 2.12.1

ymo2017

Verified signature
e87ad9c
20:47
clean rebuild: remove non-official face_detector.py

ymo2017

Verified signature
63208ba
20:47
clean rebuild: official scripts/inference.py

ymo2017

Verified signature
a9d363b
20:41
clean rebuild: restore official mask.png

ymo2017

Verified signature
dffc08e
20:41
clean rebuild: restore official latentsync/utils/image_processor.py

ymo2017

Verified signature
b574335
20:41
clean rebuild: restore official latentsync/pipelines/lipsync_pipeline.py

ymo2017

Verified signature
3443324
20:41
clean rebuild: restore official latentsync/models/unet.py

ymo2017

Verified signature
c15a947
20:41
clean rebuild: official requirements + spaces

ymo2017

Verified signature
1f9d19d
20:41
clean rebuild: pure official code, Space wrapper only

ymo2017

Verified signature
eb37d92
20:31
fix: restore cv2 sigmaX=12 feathering in paste_back (10/9 version)

ymo2017

Verified signature
f7fe71b
20:31
fix: restore LANCZOS4 for mask upscale (10/9 version)

ymo2017

Verified signature
9c0e4eb
20:31
fix: restore 10/9 shrunk mask (65% smaller)

ymo2017

Verified signature
bcaf413
20:26
fix: soften mask edge in paste_back (15x15 blur) to hide gray boundary

ymo2017

Verified signature
108ad66
20:19
fix: restore 10/9 key remap (attn2/norm2 -> audio_cross_attn, 112 keys were silently dropped -> blue squares)

ymo2017

Verified signature
a51ebd1
20:08
revert: back to 68pt landmarks

ymo2017

Verified signature
5e550ef
20:08
revert: back to face-alignment 68pt (insightface 106pt causes blue squares)

ymo2017

Verified signature
d92015b
20:02
test: remove feathering to isolate blue-square cause

ymo2017

Verified signature
0555dd5
19:56
fix: hard INTER_AREA mask (LANCZOS4 soft mask causes blue squares)

ymo2017

Verified signature
7632eae
19:23
fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

ymo2017

Verified signature
fd473e5
19:22
fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

ymo2017

Verified signature
d91dfb0
19:22
fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

ymo2017

Verified signature
8f19024
19:22
fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

ymo2017

Verified signature
eab3c0b
00:47
Fix mask edge feathering: use INTER_LANCZOS4 (as in 1.0) instead of INTER_AREA when upscaling mask.png

ymo2017

Verified signature
cbc42be

Oct 9

23:52
收小 mask：嘴部生成窗口缩小约65%，减少口罩框可见面积

ymo2017

Verified signature
cc73f87
21:05
Unconditional mask feathering + shape diag (remove dim==4 guard)

ymo2017

Verified signature
0b8667e
20:55
Feather mask edges before final pixel-space blend (sigmaX=12) to remove visible rectangular boundary

ymo2017

Verified signature
23b35de
20:44
fix: also remap ckpt resnet audio_cross_attn.norm -> model norm2 (44 resnet keys were still dropped -> color veil)

ymo2017

Verified signature
3f280f1
20:37
fix: remap checkpoint attn2/norm2 keys to audio_cross_attn (112 audio cross-attn weights were silently dropped by strict=False -> color veil)

ymo2017

Verified signature
2359c17
20:30
diag: log weight missing/unexpected keys on checkpoint load

ymo2017

Verified signature
203ec05
20:23
Replace shimmed VersatileAttention with official ByteDance SDPA forward (fixes blue-veil artifact)

ymo2017

Verified signature
cb85e10
20:02
Remove temporary debug traceback instrumentation (Space verified working)

ymo2017

Verified signature
c609701
19:56
Fix vendored _attention: restore legacy heads merge-back (batch_to_head_dim) before to_out

ymo2017

Verified signature
8366fa5
19:48
Guard legacy attention privates for diffusers>=0.29 (fix VersatileAttention AttributeError); vendored _attention shim for SDPA path

ymo2017

Verified signature
4622ff9
19:41
TEMPORARY debug: wrap generate_lip_sync_video in try/except, surface full traceback in new Textbox output

ymo2017

Verified signature
c3d85c0
19:34
Fix diffusers>=0.29 API: add reshape_heads_to_batch_dim/batch_dim_to_heads shim methods to VersatileAttention (aliases for head_to_batch_dim/batch_to_head_dim)

ymo2017

Verified signature
0a3268a
19:26
Remove xformers block: LatentSync BasicTransformerBlock's (self, valid) signature breaks every diffusers xformers setter; rely on SDPA default

ymo2017

Verified signature
0cad54e
19:20
fix: version-safe xformers call in _load_pipeline (TypeError on diffusers override)

ymo2017

Verified signature
1cbac6b
19:14
Cache pipeline per GPU worker (lazy load once); fixes 450s budget kill on cold start

ymo2017

Verified signature
26d9e8e
18:36
restore duration=300 (correct for PRO callers; 512px inference needs ~4-5 min)

ymo2017

Verified signature
10e1cd4
18:34
temp: duration 180 for anonymous smoke test (will restore 300 after)

ymo2017

Verified signature
4726b6d

Commit History

Oct 10

add missing stage2.yaml for v1.5 Running

switch to v1.5 256px (fewer mask artifacts)

fix: deploy official affine_transform.py (AlignRestore resolution)

fix: add missing enable_deepcache arg

clean rebuild: official audio2feature.py

fix: restore official face_detector.py

fix: add missing check_ffmpeg_installed to util.py

fix: unpin torch/torchvision for ZeroGPU

fix: loosen huggingface-hub for ZeroGPU gradio compat

fix: remove gradio pin to avoid ZeroGPU conflict

fix: use ZeroGPU-compatible torch 2.12.1

clean rebuild: remove non-official face_detector.py

clean rebuild: official scripts/inference.py

clean rebuild: restore official mask.png

clean rebuild: restore official latentsync/utils/image_processor.py

clean rebuild: restore official latentsync/pipelines/lipsync_pipeline.py

clean rebuild: restore official latentsync/models/unet.py

clean rebuild: official requirements + spaces

clean rebuild: pure official code, Space wrapper only

fix: restore cv2 sigmaX=12 feathering in paste_back (10/9 version)

fix: restore LANCZOS4 for mask upscale (10/9 version)

fix: restore 10/9 shrunk mask (65% smaller)

fix: soften mask edge in paste_back (15x15 blur) to hide gray boundary

fix: restore 10/9 key remap (attn2/norm2 -> audio_cross_attn, 112 keys were silently dropped -> blue squares)

revert: back to 68pt landmarks

revert: back to face-alignment 68pt (insightface 106pt causes blue squares)

test: remove feathering to isolate blue-square cause

fix: hard INTER_AREA mask (LANCZOS4 soft mask causes blue squares)

fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

fix: use official insightface 106-landmark detector (was 68-point, caused mouth shift); LANCZOS4 mask

Fix mask edge feathering: use INTER_LANCZOS4 (as in 1.0) instead of INTER_AREA when upscaling mask.png

Oct 9

收小 mask：嘴部生成窗口缩小约65%，减少口罩框可见面积

Unconditional mask feathering + shape diag (remove dim==4 guard)

Feather mask edges before final pixel-space blend (sigmaX=12) to remove visible rectangular boundary

fix: also remap ckpt resnet audio_cross_attn.norm -> model norm2 (44 resnet keys were still dropped -> color veil)

fix: remap checkpoint attn2/norm2 keys to audio_cross_attn (112 audio cross-attn weights were silently dropped by strict=False -> color veil)

diag: log weight missing/unexpected keys on checkpoint load

Replace shimmed VersatileAttention with official ByteDance SDPA forward (fixes blue-veil artifact)

Remove temporary debug traceback instrumentation (Space verified working)

Fix vendored _attention: restore legacy heads merge-back (batch_to_head_dim) before to_out

Guard legacy attention privates for diffusers>=0.29 (fix VersatileAttention AttributeError); vendored _attention shim for SDPA path

TEMPORARY debug: wrap generate_lip_sync_video in try/except, surface full traceback in new Textbox output

Fix diffusers>=0.29 API: add reshape_heads_to_batch_dim/batch_dim_to_heads shim methods to VersatileAttention (aliases for head_to_batch_dim/batch_to_head_dim)

Remove xformers block: LatentSync BasicTransformerBlock's (self, valid) signature breaks every diffusers xformers setter; rely on SDPA default

fix: version-safe xformers call in _load_pipeline (TypeError on diffusers override)

Cache pipeline per GPU worker (lazy load once); fixes 450s budget kill on cold start

restore duration=300 (correct for PRO callers; 512px inference needs ~4-5 min)

temp: duration 180 for anonymous smoke test (will restore 300 after)