nodes/kijai-ComfyUI-WanVideoWrapper
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
1,226 Commits 8 Branches 0 Tags
1cd8df5c00e519991ecdce9cbf1f89171303f447
Commit Graph
1226 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
kijai 1cd8df5c00 Update custom_linear.py 2025-10-29 02:50:11 +02:00
kijai d2614a9a49 Merge branch 'main' into longcat 2025-10-29 02:33:37 +02:00
kijai 083a8458c4 Register lora diffs as buffers to allow them to work with block swap 
unmerged loras (non GGUF for now) will now be moved with block swap instead of always loaded from cpu to reduce device transfers and allow torch compile full graph
 2025-10-29 02:33:26 +02:00
kijai 1c2f17e8d7 Add utility node to split sampler from settings 
For cleaner previews
 2025-10-29 02:24:24 +02:00
kijai 9d45b9f0de Use comfy core Conv3D workaround for VAE rather than the fp32 cast 2025-10-29 02:23:49 +02:00
Jukka Seppänen 833c6f50c7 Merge pull request #1581 from chengzeyi/fix-ref-conv-dtype-mismatch 
Fix dtype mismatch in ref_conv forward pass
 2025-10-28 14:27:43 +02:00
chengzeyiandClaude d15cf3001f Fix dtype mismatch in ref_conv forward pass 
This commit fixes a RuntimeError that occurs when using Fun-Control
reference images: "Input type (float) and bias type (c10::Half)
should be the same"

Root cause:
- Commit 1ba1a16 changed the dtype handling strategy to convert
  the main latent `x` to `base_dtype` instead of converting
  embeddings to match `x.dtype`
- This caused `fun_ref` input to be in a different dtype than
  the `ref_conv` layer's weights and bias
- Line 2324 already handles this correctly for `attn_cond` by
  converting to `self.attn_conv_in.weight.dtype`

Solution:
- Convert `fun_ref` to match `self.ref_conv.weight.dtype` before
  passing through the convolution layer
- This follows the same pattern used for `attn_cond` on line 2324

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
 2025-10-28 12:08:23 +00:00
kijai eebbcd5ee0 Update model.py 2025-10-28 02:04:50 +02:00
kijai 2633119505 Update custom_linear.py 2025-10-28 01:55:25 +02:00
kijai 90908df260 Update model.py 2025-10-28 01:54:42 +02:00
kijai c80a488f70 Use fp32 norms for other models too and other fixes 2025-10-28 01:52:48 +02:00
kijai e69e068b57 Update gguf.py 2025-10-27 21:02:59 +02:00
kijai 54c45500b0 Apply lora diffs with unmerged loras too 2025-10-27 21:01:45 +02:00
kijai e560366600 Update model.py 2025-10-27 18:55:30 +02:00
kijai f880b321c6 Allow compile in lora application 2025-10-27 01:34:32 +02:00
kijai 51fcbd6b3d Revert "Allow compile here" 
This reverts commit f583b56878.
 2025-10-27 01:07:03 +02:00
kijai f583b56878 Allow compile here 2025-10-27 01:06:39 +02:00
kijai c59e52ca44 Precision adjustments 2025-10-27 00:23:32 +02:00
kijai a0bdf20817 Some cleanup and allow full block swap 2025-10-26 23:05:17 +02:00
kijai 8ad7e50f33 Fix cross attention split point 2025-10-26 22:07:19 +02:00
kijai d504c96174 Separate attention for input images like in original 2025-10-26 19:14:47 +02:00
kijai 43acf83adb Update model.py 2025-10-26 16:57:25 +02:00
kijai fb00932cad Init 
https://huggingface.co/Kijai/LongCat-Video_comfy/tree/main
 2025-10-26 16:36:20 +02:00
kijai d74cfc54e8 Don't zero the extra frames 
Possible fix for SVI "shot" method
 2025-10-24 15:28:35 +03:00
kijai cfa883767f Update utils.py 2025-10-24 12:50:49 +03:00
kijai 88a60d71ab Fix 2025-10-23 23:25:20 +03:00
kijai b3ad381a65 not all torch versions have this 2025-10-23 21:12:45 +03:00
kijai 41168b1e82 Support light VAE 
https://huggingface.co/lightx2v/Autoencoders/tree/main
 2025-10-23 12:31:55 +03:00
kijai 7251c996d2 Load these LoRA keys too 
Even though doesn't seem to do anything? At least silences the errors
 2025-10-22 20:43:45 +03:00
kijai 67fcf0ba52 Reduce needless torch.compile recompiles 2025-10-22 13:29:13 +03:00
kijai aa610cde2b VACE: Support per context window reference images 2025-10-22 11:24:30 +03:00
kijai 6b286552b7 MocHa: remove possibly alpha channel from inputs 2025-10-21 23:26:17 +03:00
kijai 2ea824b9ab Mocha: Fix context window slicing 2025-10-21 20:33:03 +03:00
kijai 001040abc7 Update nodes.py 2025-10-21 20:27:37 +03:00
kijai 746d46e137 WanVideoScheduler: Always show split step sigma 2025-10-21 20:14:54 +03:00
kijai 089329ef2e MoCha: revert some RoPE changes 2025-10-21 20:12:40 +03:00
kijai 5c1f64197e Check for actual scale_weights even if the model for some reason doesn't have the scaled_fp8 key 2025-10-21 19:40:50 +03:00
kijai 54e938cd70 MoCha: Fix context window mask 2025-10-21 19:40:01 +03:00
kijai 7495db7669 Update nodes_sampler.py 2025-10-21 19:15:57 +03:00
kijai 4e36aee658 MoCha: experimental context windows support 2025-10-21 18:32:40 +03:00
kijai 56120d633e Update nodes.py 2025-10-21 17:59:37 +03:00
kijai 1f0861b649 MoCha: modify RoPE function to be more torch.compile friendly 2025-10-21 17:54:25 +03:00
kijai 7916f89c33 Add updated MoCha example for lower VRAM 2025-10-21 17:27:41 +03:00
Jukka Seppänen e294f417c0 Merge pull request #1501 from unrealMJ/mocha 
[Feature] Request to add MoCha: End-to-End Video Character Replacement without Structural Guidance
 2025-10-21 17:17:21 +03:00
unrealMJ ee011e66e9 update workflow 2025-10-21 17:28:50 +08:00
unrealMJ d7fc563581 add mocha workflow 2025-10-21 10:00:51 +08:00
unrealMJ 88defbfdd1 add MoCha 2025-10-21 09:40:54 +08:00
Jukka Seppänen 74f33df658 Merge pull request #1493 from HM-RunningHub/fix-uni3c-context-options 
Fix Uni3C + Context Options compatibility (Issue #1491)
 2025-10-20 22:33:38 +03:00
wenjian 487c400e8e Fix Uni3C + Context Options compatibility issue (Issue #1491) 
Add uni3c_data parameter to predict_with_cfg() call in context windowing loop.

Without this parameter, Uni3C camera effects were not working when using
Context Options for long video generation. This makes the context windowing
behavior consistent with multitalk and wananimate sampling modes.
 2025-10-21 02:31:47 +08:00
kijai 200f6943e3 Add sageattn mode that allows torch.compile 
Latest wheel from woct0rdho includes the torch.compile fix:

https://github.com/woct0rdho/SageAttention/releases

Based on my quick testing this reduces peak VRAM usage a bit when running sageattn + torch.compile
 2025-10-20 15:16:43 +03:00