Hong Zhang
4ec4d9c20a
Merge pull request #1354 from mi804/low_vram_training_ds
...
low vram training with deepspeed zero3
2026-03-17 16:09:52 +08:00
Zhongjie Duan
7650e9381e
Update audio.py ( #1349 )
2026-03-13 17:57:14 +08:00
Hong Zhang
8c9ddc9274
support loading ltx2.3 stage2lora by statedict ( #1348 )
...
* support ltx2.3 stage2lora by statedict
* bug fix
* bug fix
2026-03-13 17:19:18 +08:00
Hong Zhang
681df93a85
Mova ( #1337 )
...
* support mova inference
* mova media_io
* add unified audio_video api & fix bug of mono audio input for ltx
* support mova train
* mova docs
* fix bug
2026-03-13 13:06:07 +08:00
Hong Zhang
4741542523
Ltx2.3 a2v& retake video and audio ( #1346 )
...
* temp commit
* support ltx2 a2v
* support ltx2.3 retake video and audio
* add news
* minor fix
2026-03-12 14:16:01 +08:00
Hong Zhang
c927062546
Merge pull request #1343 from mi804/ltx2.3_multiref
...
Ltx2.3 multiref
2026-03-10 17:31:05 +08:00
Zhongjie Duan
f3ebd6f714
Merge pull request #1342 from modelscope/ltx2-default-prompt
...
add default negative prompt of ltx2
2026-03-10 15:10:51 +08:00
Artiprocher
959471f083
add default negative prompt of ltx2
2026-03-10 15:10:03 +08:00
Hong Zhang
d9228074bd
refactor ltx2 stage2 pipeline ( #1341 )
...
* refactor ltx2 pipeline
* fix bug
2026-03-10 13:55:40 +08:00
Hong Zhang
b272253956
Ltx2.3 i2v training and sample frames with fixed fps ( #1339 )
...
* add 2.3 i2v training scripts
* add frame resampling by fixed fps
* LoadVideo: add compatibility for not fix_frame_rate
* refactor frame resampler
* minor fix
2026-03-09 20:32:02 +08:00
Hong Zhang
7bc5611fb8
ltx2.3 bugfix & ic lora ( #1336 )
...
* ltx2.3 ic lora inference&train
* temp commit
* fix first frame train-inference consistency
* minor fix
2026-03-09 16:33:19 +08:00
Artiprocher
13eff18e7d
remove unnecessary params in cache
2026-03-09 14:09:30 +08:00
mi804
d40efe897f
ltx2.3 train
2026-03-06 18:08:42 +08:00
Zhongjie Duan
c9c2561791
Merge pull request #1333 from mi804/ltx2.3
...
ltx2.3 docs
2026-03-06 16:53:56 +08:00
mi804
ed9e4374af
ltx2.3 docs
2026-03-06 16:45:12 +08:00
Zhongjie Duan
2a0eb9c383
support ltx2.3 inference ( #1332 )
2026-03-06 16:24:53 +08:00
mi804
73b13f4c86
support ltx2.3 inference
2026-03-06 16:07:17 +08:00
Zhongjie Duan
31ba103d8e
Merge pull request #1330 from modelscope/ses-doc
...
Research Tutorial Sec 2
2026-03-06 14:25:45 +08:00
Zhongjie Duan
6bcb99fd2e
Merge branch 'main' into layercontrol_v2
2026-03-03 21:04:04 +08:00
Artiprocher
add6f88324
bugfix
2026-03-03 15:33:42 +08:00
Zhongjie Duan
430b495100
Merge pull request #1321 from mi804/bugfix
...
fix qwen_text_encoder bug in transformers>=5.2.0
2026-03-03 13:02:45 +08:00
mi804
62ba8a3f2e
fix qwen_text_encoder bug in transformers>=5.2.0
2026-03-03 12:44:36 +08:00
Zhongjie Duan
237d178733
Fix LoRA compatibility issues. ( #1320 )
2026-03-03 11:08:31 +08:00
Zhongjie Duan
b3ef224042
support Anima gradient checkpointing ( #1319 )
2026-03-02 19:06:55 +08:00
Zhongjie Duan
6d671db5d2
Support Anima ( #1317 )
...
* support Anima
Co-authored-by: mi804 <1576993271@qq.com >
2026-03-02 18:49:02 +08:00
Zhongjie Duan
29cd5c7612
Merge pull request #1275 from Mr-Neutr0n/fix-dit-none-check
...
Fix AttributeError when pipe.dit is None during split training
2026-03-02 10:25:11 +08:00
Zhongjie Duan
ff4be1c7c7
Merge pull request #1293 from Mr-Neutr0n/fix/trajectory-loss-div-by-zero
...
fix: prevent division by zero in TrajectoryImitationLoss at final denoising step
2026-03-02 10:21:39 +08:00
Zhongjie Duan
6b0fb1601f
Merge pull request #1296 from Explorer-Dong/fix/wan_vae
...
fix: WanVAE2.2 encode and decode error
2026-03-02 10:19:36 +08:00
mi804
1a380a6b62
minor fix
2026-02-28 11:09:10 +08:00
mi804
8b9a094c1b
ltx iclora train
2026-02-27 18:43:53 +08:00
mi804
5996c2b068
support inference
2026-02-27 16:48:16 +08:00
mi804
a18966c300
support ltx2 gradient_checkpointing
2026-02-26 19:19:59 +08:00
mi804
f48662e863
update docs
2026-02-26 11:10:00 +08:00
mi804
8d8bfc7f54
minor fix
2026-02-25 19:04:10 +08:00
mi804
8e15dcd289
support ltx2 train -2
2026-02-25 18:06:02 +08:00
mi804
586ac9d8a6
support ltx-2 training
2026-02-25 17:19:57 +08:00
mi804
ee73a29885
qwen_image layercontrol v2
2026-02-24 15:19:16 +08:00
Mr_Dwj
fc11fd4297
chore: remove invalid comment code
2026-02-13 09:38:14 +08:00
Mr_Dwj
bd3c5822a1
fix: WanVAE2.2 decode error
2026-02-13 01:13:08 +08:00
Mr_Dwj
96fb0f3afe
fix: unpack Resample38 output
2026-02-12 23:51:56 +08:00
Mr-Neutr0n
b68663426f
fix: preserve sign of denominator in clamp to avoid inverting gradient direction
...
The previous .clamp(min=1e-6) on (sigma_ - sigma) flips the sign when
the denominator is negative (which is the typical case since sigmas
decrease monotonically). This would invert the target and cause
training divergence.
Use torch.sign(denom) * torch.clamp(denom.abs(), min=1e-6) instead,
which prevents division by zero while preserving the correct sign.
2026-02-11 21:04:55 +05:30
Mr-Neutr0n
0e6976a0ae
fix: prevent division by zero in trajectory imitation loss at last step
2026-02-11 19:51:25 +05:30
Hong Zhang
b3b63fef3e
Add readthedocs for diffsynth-studio
...
* add conf docs
* add conf docs
* add index
* add index
* update ref
* test root
* add en
* test relative
* redirect relative
* add document
* test_document
* test_document
2026-02-10 19:51:04 +08:00
Artiprocher
fddc98ff16
fix mix-precision issues in low-version torch
2026-02-10 11:12:50 +08:00
Zhongjie Duan
dc94614c80
Merge pull request #1256 from Feng0w0/npu_fused
...
[model][NPU]:Add NPU fusion operator patch to Zimage model to improve performance
2026-02-09 20:08:44 +08:00
feng0w0
e56a4d5730
[model][NPU]:Add NPU fusion operator patch to Zimage model to improve performance
2026-02-09 12:31:34 +08:00
feng0w0
3f8468893a
[model][NPU]:Add NPU fusion operator patch to Zimage model to improve performance
2026-02-09 09:51:06 +08:00
Mr-Neutr0n
6383ec358c
Fix AttributeError when pipe.dit is None
...
When using split training with 'sft:data_process' task, the DiT model
is not loaded but the attribute 'dit' exists with value None. The
existing hasattr check returns True but then accessing siglip_embedder
fails.
Add an explicit None check before accessing pipe.dit.siglip_embedder.
Fixes #1246
2026-02-07 05:23:11 +05:30
Zhongjie Duan
1b47e1dc22
Merge pull request #1272 from modelscope/zero3-fix
...
Support DeepSpeed ZeRO 3
2026-02-06 16:33:12 +08:00
Artiprocher
b0bf78e915
refine code & doc
2026-02-06 16:27:23 +08:00
