search

1038lab / ComfyUI-OmniGen

ComfyUI-OmniGen - A ComfyUI custom node implementation of OmniGen, a powerful text-to-image generation and editing model.
MIT License
104 stars 10 forks source link

[Feature Request] Add tiled VAE and option to disable K/V cache #20

Open set-soft opened 1 week ago

set-soft commented 1 week ago

I guess GGUF is the real solution, but you should add a tiled VAE retry, like in the default VAE decoder of Comfy_UI. BTW: Can you just decouple the VAE decode and take advantage of the automatic tiled VAE? Is really annoying to waste 10 minutes to get a fail at the end because the VAE decode didn't have enough memory.

Another think that can help: https://github.com/chflame163/ComfyUI_OmniGen_Wrapper Adds an option to disable the K/V cache, it is slower, but reduces the VRAM needed.

1038lab commented 1 week ago

Currently, this model utilizes the SDXL VAE, and I'm unable to modify the underlying structure. Regarding the suggestion to disable the K/V cache, it’s worth noting that under low VRAM conditions, this model already operates slowly, and I'm not sure users would want to invest additional time for image generation.

However, if there’s significant interest in the option to disable the K/V cache, please let me know, and I’ll consider creating a branch for this feature. Your feedback is valuable!