Go to file

Yoshimasa Niwa 5ca4bbf319 Workaround pad problem on mps

When using `torch.nn.functional.pad` with tensor that size is
larger than 2^16 (65526), the output tensor would be broken.

This patch moves tensor to CPU to workaround the problem.
It doesn't too much impacts in terms of speed of vea on mps.

2024-11-05 12:46:13 +09:00

examples

make compatible with comfy cliptextencode

2024-10-28 12:23:14 +02:00

mochi_preview

Workaround pad problem on mps

2024-11-05 12:46:13 +09:00

__init__.py

initial commit

2024-10-23 15:34:22 +03:00

.gitattributes

Initial commit

2024-10-23 11:58:29 +03:00

.gitignore

cleanup

2024-10-23 15:45:20 +03:00

fp8_optimization.py

Add alternative VAE decoding node

2024-10-25 15:30:20 +03:00

infer.py

initial commit

2024-10-23 15:34:22 +03:00

latent_preview.py

Add sampler preview

2024-11-01 06:18:41 +02:00

LICENSE

Create LICENSE

2024-10-25 01:33:30 +03:00

mz_gguf_loader.py

Update mz_gguf_loader.py

2024-10-28 04:07:11 +02:00

nodes.py

Update nodes.py

2024-11-04 15:10:15 +02:00

readme.md

Update readme.md

2024-10-24 02:55:09 +03:00

requirements.txt

Add accelerate

2024-10-23 16:05:19 +03:00

readme.md

ComfyUI wrapper nodes for Mochi video generator

WORK IN PROGRESS

https://github.com/user-attachments/assets/a714b70f-dcdb-4f91-8a3d-8da679a28d6e

Can use flash_attn, pytorch attention (sdpa) or sage attention, sage being fastest.

Depending on frame count can fit under 20GB, VAE decoding is heavy and there is experimental tiled decoder (taken from CogVideoX -diffusers code) which allows higher frame counts, so far highest I've done is 97 with the default tile size 2x2 grid.

Models:

https://huggingface.co/Kijai/Mochi_preview_comfy/tree/main

model to: ComfyUI/models/diffusion_models/mochi

vae to: ComfyUI/models/vae/mochi

There is autodownload node (also will be normal loader node)