GGUf version of H3 Eros Max
Description
FAQ
Comments (12)
Is this the TURBO-hybrid 8 step version?
Yes, this is the version https://civitai.red/models/2851079/h3-eros-max?modelVersionId=3275279 it works in 6-8 steps.
Q8 or in8, which one is faster on 30xx gpus?
I recommend this model. Excellent speed with good quality. Unfortunately, the version is old :( https://huggingface.co/taxexempt/Custom-MiniMaX-H3-mixed-quants/blob/main/diffusion_models/h3ErosMax_beta3__bf16_w4a8_int8_convrot_mix.safetensors
I also recommend the accelerator, with an acceleration of about 50% on my 4080 and sage attention. https://github.com/StanLukuvka/ComfyUI-MiniMax-H3-SPEED
int8 is 10-15 % slower on my 3060 12gb
if you can fit the normal version, non-gguf, use the normal version... This should only be used if you get oof out of memory with the normal int8 version.
I don't understand which node loader I should use for this model. All my old nodes throw an error when loading GGUF; they don't recognize the architecture for minimax h3.
same here
I believe this is how I got it working
Most likely, you need to update comfy. I use this because it loads using dynamic vram. https://github.com/molbal/ComfyUI-GGUF
Unet Loader (Dynamic VRAM) by gguf-reboot