Seperate files instead of checkpoints.

#11
by natalie5 - opened

The current model files include vocoder, VAE, Audio VAE alongisde the transformer blocks itself. This is an extremely unoptimized way of loading the model when being used with low VRAM. Having seperate VAE, vocoder files will make implementations/quantizations, VRAM use much better. This also leads to duplicated files where each checkpoint has the same VAEs, vocoder, it will save 4GB of RAM when sampling, and ~14GB in SSD space.

Looks like Kijai is also adding some versions ;-)
https://huggingface.co/Kijai/LTXV2_comfy

@RuneXX those newly added files are most likely for diffusers, not comfyui. The commit message that adds them says that they are weights for diffusers, as well.

@RuneXX those newly added files are most likely for diffusers, not comfyui. The commit message that adds them says that they are weights for diffusers, as well.

https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn these work in native comfy, we just need the diffusion model itself (without all the extra fluff) to make it fully seperate, plus the audio vae needs some code changes to make it work with the normal vae loader

image

Sign up or log in to comment