kijai
04534e5eed
Allow using multiple VACE embeds, fix VACE blockswap error when vace blocks not set and blockswap is set
...
Experimental
2025-04-04 15:23:39 +03:00
kijai
e4912d03de
non_blocking teacache offload
2025-04-03 20:06:15 +03:00
kijai
1aa6498bbe
Compile VACE blocks as well
2025-04-03 17:05:39 +03:00
kijai
4cebb55032
more robust vace check
2025-04-02 18:41:39 +03:00
kijai
390bc2a6cb
Update model.py
2025-04-02 18:38:16 +03:00
kijai
c33ba8b05c
Update model.py
2025-04-02 18:27:40 +03:00
kijai
f820f38c8f
Allow offloading and block swapping VACE blocks as well
2025-04-02 17:27:17 +03:00
kijai
0801394cb5
Fix TeaCache memory issue
2025-04-02 16:14:24 +03:00
kijai
ce6522c227
Add start/end percent controls for VACE
2025-04-02 11:10:12 +03:00
kijai
96b236c4df
Basic VACE (preview version) support
...
no ref image or masking yet, but control signal seems to work
2025-04-01 22:05:33 +03:00
kijai
3869b0482b
cleanup
2025-04-01 19:13:53 +03:00
kijai
23c9a6d634
Add dynamo_recompile_limit as adjustable option, revert to old euler for flowedit
2025-04-01 11:44:45 +03:00
kijai
104c632359
some cleanup
2025-03-28 01:58:47 +02:00
kijai
2d56d69ccb
Add alternative TeaCache mode, better frame matching for control workflow
...
Based on the coefficients from here:
https://github.com/ali-vilab/TeaCache/pull/51
2025-03-27 19:44:37 +02:00
kijai
921e577ad9
Update model.py
2025-03-26 18:59:13 +02:00
kijai
e5df47a2d3
Support Fun-InP control T2V
2025-03-26 18:36:08 +02:00
kijai
0e0962d6e8
support 1.3B Fun Control -model
2025-03-26 16:39:27 +02:00
kijai
a1938107cd
Allow splitting self attention as well when using multiple prompts
2025-03-25 21:05:08 +02:00
kijai
b4dc58fd9f
Split cross attention when using multiple prompts and/or clip image embeds
...
Allows for (pretty weak) prompt travel and may help with the endframe workflow.
2025-03-25 00:07:41 +02:00
kijai
4b8672ac3b
Update model.py
2025-03-23 02:39:34 +02:00
kijai
8fbf92ae93
Improved loop stitching on decode
2025-03-22 23:05:32 +02:00
kijai
309240c0f4
fix T2V (again)
2025-03-22 10:14:07 +02:00
kijai
2863956da7
workflow updates
2025-03-22 02:45:11 +02:00
kijai
b415fe78f1
fix T2V
2025-03-21 20:44:54 +02:00
kijai
c3e96fd6f9
Experimental clip embed enhancement through tiling
...
Allows more detail in the clip embeds as well as details from non square images without stretching
Idea and most of the code is from Matteo:
https://github.com/cubiq/ComfyUI_IPAdapter_plus/blob/9d076a3df0d2763cef5510ec5ab807f6632c39f5/utils.py#L181
2025-03-21 19:41:49 +02:00
kijai
173aff2367
fix I2V flowedit and maybe allow compile with vram management
2025-03-16 03:00:14 +02:00
kijai
81a08d0e59
Make non_blocking offloading optional for low RAM usecases
2025-03-16 01:48:45 +02:00
kijai
d455cc89a3
vram management fix
2025-03-15 16:40:39 +02:00
kijai
58e5737032
Add option to use comfy native RoPE calculations
...
comfy rope doesn't use complex numbers and can be torch.compiled, then it's faster and uses less memory
2025-03-15 15:45:53 +02:00
kijai
ff4bb8085e
remove print
2025-03-13 18:17:46 +02:00
kijai
8ac0da07c6
Add SLG
...
Credits to AmericanPresidentJimmyCarter: https://github.com/deepbeepmeep/Wan2GP/pull/61
2025-03-13 17:29:15 +02:00
kijai
f07850cf9c
fix vram management
2025-03-13 02:41:55 +02:00
kijai
71fd6c7208
rope tweaks
2025-03-13 02:26:40 +02:00
kijai
cd9e9dc891
Update model.py
2025-03-12 19:58:53 +02:00
kijai
5c63ff5bce
Faster RoPE
2025-03-12 18:36:28 +02:00
kijai and spacepxl
ffaf1afe4e
Support control-lora
...
https: //huggingface.co/spacepxl/Wan2.1-control-loras/blob/main/wan2.1-1.3b-control-lora-tile-v0.1_comfy.safetensors
Co-Authored-By: spacepxl <143970342+spacepxl@users.noreply.github.com >
2025-03-10 21:31:06 +02:00
kijai
07848ab0b5
fix
2025-03-09 23:26:46 +02:00
kijai
5a68948607
enhance-a-video tweak
2025-03-09 18:42:06 +02:00
kijai
389bbd343c
Update TeaCache docs
2025-03-08 12:13:52 +02:00
kijai
32465c2787
TeaCache tweaks for context windowing and FlowEdit
2025-03-05 21:02:34 +02:00
kijai
de3e0f9ef2
update TeaCache
2025-03-05 13:35:18 +02:00
kijai
c4472c2906
Add FlowEdit and some code refactors
2025-03-05 00:54:08 +02:00
kijai
0183a03bf9
revert and fix again...
...
I'll clean this later
2025-03-03 01:38:51 +02:00
kijai
24fb1b42fe
bug fixes
2025-03-03 01:35:36 +02:00
kijai
d3129c6996
Add alternative offloading method from Diffsynth
2025-03-02 14:37:08 +02:00
kijai
ac98a0318b
Update model.py
2025-03-02 10:52:25 +02:00
kijai
55dcc33d00
Add block swap debug prints and cleanup unused code
2025-03-02 10:49:46 +02:00
kijai
9087791b75
Allow setting TeaCache device to avoid the VRAM usage increase
2025-03-02 09:57:02 +02:00
zhuhz22
b366a40ab6
fix: riflex
2025-03-02 14:38:08 +08:00
kijai
fbb432ce0f
Experimental TeaCache
...
I haven't figured out the whole coefficiency calculation part, however I noticed it's already super close to the time embed as it is, if we just skip the initial steps from the calculations. Seems to work okayish with the 1.3B model at least.
2025-03-02 01:34:55 +02:00
kijai