blazing fast, only 1 minute 10 seconds generating 5 second clip in 0.4 mp resolution.
NEW using Ultimate upscale 1.0mp resolution or 768p got 3 minutes in my rtx 4060ti 16gb
UPDATE!!!:
i ditch sage attention and use Comfy kitchen + sol attention instead, but i'm still use kjnodes. here is the link to install custom nodes:
kjnodes: https://github.com/kijai/ComfyUI-KJNodes
Comfy Kitchen: https://github.com/Comfy-Org/comfy-kitchen
Sol attention Triton: https://github.com/kijai/ComfyUI-SolAttn_triton
use this model from huggingface: https://huggingface.co/MATLOWAI/minimax-h3-fused-turbo-int8-convrot/tree/main/diffusion_models
if you like my work and want to support me, you can donate me to my ko-fi: https://ko-fi.com/megaaziib
get minimax ultimate upscale: https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale
for upscale model, you need to set the second resolution either 0.6, 0.7- 1.0. megapixel. if you get oom or generating too long, try to decrease resolution to 0.9 or 0.8 mp.
change manualsigmas according your preference steps:
4 steps:
0.9999166, 0.9728326, 0.9230769, 0.8, 0.0
6 steps:
0.9999166, 0.9868421, 0.9638554, 0.9230769, 0.8695652, 0.8, 0.0
8 steps:
1.0, 0.9882352941, 0.9729729730, 0.9523809524, 0.9230769231, 0.8780487805, 0.8, 0.6315789474, 0.0
Description
use comfy kitchen + sol attention
FAQ
Comments (13)
I see the prompt format you have used is different from the one provided by the Minimax H3 prompt guide. Why?
i just make quick testing with chatgpt, you can change to normal prompt from minimax h3 documentation instead.
@aziib Ok. Thanks for sharing the workflows. Really appreciate it.
This works ok. I got it to do image to video but had to use a lightning lora to make up for the too few steps it does. Seems the resolution table is slightly different. The animation seems a little weaker than my old setup but it's definitely faster.
the model still in development and not good enough, i switch to this instead: https://huggingface.co/MATLOWAI/minimax-h3-fused-turbo-int8-convrot/tree/main/diffusion_models
Useless for any practice quality.
Incompatible with latent upscale - slower than int8 convrot with turbo lora.
No i2v and ref2v.
You are using the fast model and the turbo lora together for i2v?
damn you are right, i forgot to remove it when testing.
I get a 'unet unexpected' error when generating with the Fastvideo_vsa_datafree_1300step_4step_int8_convrot
Comfyui is current and working in every other way... what might be missing for the FastVideo checkpoint to run please?
i dont use that model anymore, i prefer this model to have better quality and same speed: https://huggingface.co/MATLOWAI/minimax-h3-fused-turbo-int8-convrot/tree/main/diffusion_models
Yes - already done that and the WF works really well, thanks. But I'd like to find the answer so I can do my own testing with the Fast model
@aziib compared of kijai version which better
minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors
@roxgame i tried kijai version, it still need turbo lora, not good compare MATLOWA has already turbo inside it.