kijai
f86d2b84a0
Update nodes.py
2025-07-27 15:34:57 +03:00
kijai
3567202e75
Fix compile and support radial attn in the DF sampler
2025-07-27 14:19:31 +03:00
Jukka Seppänen
c0a4e8aa87
Update readme.md
2025-07-27 01:43:32 +03:00
kijai
bc68d35e5a
Create wanvideo_1_3B_EchoShot_example.json
2025-07-26 19:25:04 +03:00
kijai
cb36e1013e
Fix
2025-07-26 19:01:58 +03:00
kijai
df1476bf42
Support EchoShot
...
https://github.com/D2I-ai/EchoShot
2025-07-26 18:51:29 +03:00
kijai
9a1ab1c656
only show lora metadata if it actually exists
2025-07-26 14:32:53 +03:00
kijai
5c36a4b111
minor memory tweaks
2025-07-26 14:19:07 +03:00
kijai
3157e0389e
Update nodes.py
2025-07-26 11:50:24 +03:00
kijai
34c2cc23a9
Update nodes_model_loading.py
2025-07-25 20:49:41 +03:00
kijai
a35eb7dfa4
Basic sageattn3 support
...
Probably going to need further timestep scheduling as the quality of sage3 on it's own is pretty bad.
2025-07-25 20:48:34 +03:00
kijai
76750eec05
Allow loading aitoolkit/lycoris format
2025-07-25 17:24:43 +03:00
kijai
b89cf970e1
Fix GGUF + LoRA + VACE
...
Always something part two...
2025-07-25 16:44:43 +03:00
kijai
794a38cc7e
Further GGUF+LoRA fixes
2025-07-25 00:40:48 +03:00
kijai
73b7444baa
remove print
2025-07-24 23:54:30 +03:00
kijai
ea36decf07
Fix GGUF + LoRA + torch.compile
...
Always something...
2025-07-24 23:52:18 +03:00
kijai
b3820055d0
Fix GGUF with LoRAs and support GGUF witgh SetLoras -node
2025-07-24 14:59:59 +03:00
kijai
35e637cbfd
Allow merging LoRA to fp8 scaled models
2025-07-24 11:51:36 +03:00
kijai
d6425cab02
Update nodes_model_loading.py
2025-07-24 00:08:39 +03:00
kijai
59cd0dc272
Update nodes.py
2025-07-23 18:11:15 +03:00
kijai
838803d053
Update nodes_model_loading.py
2025-07-23 17:58:28 +03:00
kijai
bdce322e65
remove loras if SetLora bypassed/disconnected
2025-07-23 17:51:59 +03:00
kijai
66226a8a1b
why was this ever this way...
2025-07-23 17:17:27 +03:00
kijai
3ed58330d9
better
2025-07-23 16:56:33 +03:00
kijai
65ac9fa6d9
cache positive and negative separately, add cache to single encode node as well
2025-07-23 16:33:11 +03:00
kijai
5e56e3649f
fix control lora
2025-07-23 16:20:30 +03:00
kijai
0212ad7a2d
cache only the prompts
...
so we can disconnect T5 completely if text embed exists in cache
2025-07-23 15:33:32 +03:00
kijai
1a1af02912
Allow using different init image for context windows beyond first one
...
This can be useful with models like MAGREF that behave differently when the init image is padded with white, idea is to use full init for first window and padded image for the rest, so that it works more like reference and doesn't force the window to snap back to the init.
2025-07-23 12:51:57 +03:00
kijai
9d9b188b0b
Update fp8_optimization.py
2025-07-23 01:22:42 +03:00
kijai
fa93f5b3c1
Add optional prompt disk caching
...
Caches text embeds to disk with unique hash based on T5 name, dtype and prompts. If found on disk skips encoding, persists after restarts.
2025-07-23 00:52:56 +03:00
kijai
729b6fdf7c
support sparse sageattn2 for radial attn
...
Needs this installed:
https://github.com/Radioheading/Block-Sparse-SageAttention-2.0
2025-07-23 00:27:14 +03:00
kijai
9c45f3a97d
Update nodes_model_loading.py
2025-07-23 00:10:14 +03:00
kijai
eea896dace
Update nodes_model_loading.py
2025-07-22 18:25:03 +03:00
kijai
70fcdff3c5
Update nodes_model_loading.py
2025-07-22 17:36:41 +03:00
kijai
bfd6af8141
Update nodes_model_loading.py
2025-07-22 15:41:32 +03:00
kijai
98fa84bb17
Update nodes_model_loading.py
2025-07-22 13:58:43 +03:00
kijai
3988accdf3
Update nodes_model_loading.py
2025-07-21 23:44:24 +03:00
kijai
9911b5e6f0
oops
2025-07-21 22:49:05 +03:00
kijai
91515717d6
bump version
2025-07-21 22:46:50 +03:00
kijai
296baa30ce
Add WanVideoSetLoRAs
...
Node to set the LoRA weights to use with the unmerged LoRA mode, not able to merge LoRAs but allows instant LoRA switching without any loading times. The effect of unmerged LoRAs is stronger and differs from merged LoRAs.
2025-07-21 22:39:17 +03:00
kijai
41d8bd9ec9
fix for context windows
2025-07-21 20:09:53 +03:00
kijai
4bf78e5316
Possibly reduce LoRA loading memory usage
2025-07-21 15:25:07 +03:00
kijai
6ef9224dfe
Update model.py
2025-07-21 13:51:33 +03:00
kijai
aeac12ed7a
Only convert layers that actually have scaled weights...
2025-07-21 01:57:27 +03:00
kijai
29ce253bb6
Support fp8_scaled models and allow running LoRAs unmerged on other models as well
2025-07-21 01:37:30 +03:00
kijai
59add151eb
Update attention.py
2025-07-20 16:22:27 +03:00
kijai
e5b125cd89
Allow setting any block as dense for Radial attention
...
Also added WanVideoBlockList helper to to create list of ints that can be inserted to the dense_blocks -input.
2025-07-20 16:21:37 +03:00
kijai
0dba876979
Correct radial attn mask size when using context windows
2025-07-20 13:22:56 +03:00
kijai
bb82d63453
Update nodes.py
2025-07-20 13:08:40 +03:00
kijai
768a4b2d90
Fix radial attention cache not resetting on resolution change, add more robust supported dimension check
2025-07-20 12:15:01 +03:00
kijai
9f75b7e051
Fix tiled_vae + endframe
2025-07-20 02:33:24 +03:00
kijai
1f25bba463
Merge branch 'pr/828'
2025-07-20 02:13:26 +03:00
kijai
47978275bb
Update nodes.py
2025-07-20 02:13:07 +03:00
kijai
edd9b20691
Add res_multistep
2025-07-20 01:41:17 +03:00
kijai
0f6fe96626
Update __init__.py
2025-07-18 23:01:03 +03:00
kijai
feddafcea2
Update model.py
2025-07-18 18:52:32 +03:00
kijai
5a305a7f9f
include sparse_sage_API code
...
https://github.com/jt-zhang/Sparse_SageAttention_API
2025-07-18 17:16:43 +03:00
kijai
cf18ecc2da
update i2v basic example
2025-07-18 16:42:02 +03:00
kijai
c68f5f7439
move node
2025-07-18 16:41:49 +03:00
kijai
c08296747f
more refactoring
2025-07-18 16:31:34 +03:00
kijai
6b9565ed63
Code refactoring
2025-07-18 16:15:50 +03:00
kijai
8b8da7d8fe
Update nodes.py
2025-07-18 14:41:38 +03:00
kijai
8538c517e6
Merge branch 'radial_attn'
2025-07-18 14:40:10 +03:00
kijai
ed11e2bb3e
Update nodes.py
2025-07-18 14:37:11 +03:00
kijai
a1cc320c36
force using the set node
2025-07-18 14:26:45 +03:00
kijai
0e920e5b18
support VACE
2025-07-18 14:04:04 +03:00
kijai
55fc8a13e0
better mask cache, optimize
2025-07-18 13:55:06 +03:00
kijai
8919e65bab
cache the mask
2025-07-18 01:43:45 +03:00
kijai
a7166fc1a4
fix and rename args to be clearer
2025-07-18 01:28:01 +03:00
kijai
712eecf64b
more compile friendly
2025-07-18 00:53:12 +03:00
kijai
605011a237
allow setting dense attention mode
2025-07-17 23:47:08 +03:00
kijai
9039d95721
cleanup and optimize
2025-07-17 22:40:26 +03:00
kijai
7c8020f8e8
works but slow
2025-07-17 21:49:14 +03:00
kijai
0417c8ea00
Update nodes.py
2025-07-17 20:48:16 +03:00
kijai
6ee7ad508e
init (not working)
2025-07-17 20:47:02 +03:00
Jukka Seppänen
31b2c686cf
Merge pull request #768 from Eikwang/fixunianimate
...
fix unianimate erro : OverflowError: cannot convert float infinity to…
2025-07-17 19:33:37 +03:00
kijai
e011d031a8
Merge branch 'pr/799'
2025-07-17 19:33:10 +03:00
kijai
6e5b493b28
Create wanvideo_14B_pusa_I2V_example_01.json
2025-07-17 19:27:32 +03:00
kijai
8d7504144d
Update nodes.py
2025-07-17 19:25:03 +03:00
kijai
d8e90a94c1
more advanced input for Pusa
2025-07-17 19:10:24 +03:00
kijai
6bc53b771d
refactor scheduler import
2025-07-17 15:56:55 +03:00
kijai
2263d02a0a
basic Pusa support
2025-07-17 10:06:01 +03:00
kijai
17d48e3e45
rename VACE Model - VACE Module for clarity
2025-07-14 17:38:08 +03:00
kijai
a37c0e6ac8
Update nodes.py
2025-07-14 17:28:17 +03:00
kijai
24d3de2126
Add EasyCache
2025-07-14 17:27:10 +03:00
kijai
e1f5d185ee
Use original Multitalk attention code with exactly 2 speakers
2025-07-14 17:06:48 +03:00
kijai
f96128bf2c
Make multitalk sampling actually use the seed...
2025-07-14 09:20:03 +03:00
kijai
8fd62b3e78
Multitalk wav2vec model doesn't actually need the .pt file
...
chinese-wav2vec2-base-fairseq-ckpt.pt can be safely deleted and won't be autodownloaded in the future
2025-07-14 09:13:41 +03:00
kijai
08669de287
Fix FantasyTalking when wav2vec loaded on offload_device
2025-07-13 23:52:19 +03:00
kijai
b6d156afa5
Update nodes.py
2025-07-13 19:13:20 +03:00
kijai
1e4ed578ba
Update pyproject.toml
2025-07-13 18:42:17 +03:00
kijai
365a1ce1d6
more chunking and refactoring
2025-07-12 20:18:52 +03:00
kijai
895c92febd
Update model.py
2025-07-12 03:04:46 +03:00
kijai
039b211ef7
cleanup
2025-07-12 02:06:02 +03:00
kijai
127bfa2c99
refactor self attention
2025-07-11 23:00:59 +03:00
kijai
880194a1f9
bump version: 1.2.3
2025-07-11 01:13:25 +03:00
kijai
98395f94ce
Update nodes_model_loading.py
2025-07-11 01:11:43 +03:00
kijai
1077d2323c
fix max res selection
2025-07-10 23:41:54 +03:00
kijai
15f0e62e5f
Add chunked RoPE option to reduce peak VRAM usage when not using torch.compile
2025-07-10 17:15:06 +03:00
kijai
49de5335c0
fix multi lora loader
2025-07-10 16:30:20 +03:00
kijai