TaoMate-H3-3step Lora converted to ComyUI.
Generate a MiniMax H3 video in only 3 steps.
Workflow included in video.
scheduler: simple
steps: 3
cfg: 1.0
sampler: euler
video_shift: 12 or 8
audio_shift: 3
credits: https://github.com/TaoLiveAIGC/TaoMate-H3
Converted to ComfyUI format: CZMartin22
Description
FAQ
Comments (12)
Seems another one was extracted here, a bit larger:
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/tree/main/experimental
I guess the difference is : 1.24 GB - fp16, 2.48 GB - 32.
@Mr_Einhorn KJ also released on that was 300mb, lower rank, haven't tried it yet
did you try more steps? and how's the results?
Summary of early community impressions:
The ComfyUI version seems to have a tendency toward slow-motion output, but the overall quality looks good.
Despite being trained for FL2VA, it may be worth trying it with Ref2VA as well. Strength 0.7 with 8 steps seems to give decent visual quality, although I can't really comment on the audio yet. Strength 1.0 tends to burn out the image, while 3 steps are basically useless for Ref2VA.
I also tried 4 steps at low resolution -> upscale -> 2 additional steps at high resolution. This gave me a very fast workflow with results I'm quite happy with.
Personally, I think it works particularly well for scenes that don't require a lot of motion, or as a refinement/upscaling pass.
I think we'll have a much clearer picture once the community has had more time to test it.
its taomate version the replacement of the original one or, llike a kind of wan 2.2 vs 2.1
The ComfyUI version seems to have a tendency toward slow-motion output, but the overall quality looks good.
Despite being trained for FL2VA, it may be worth trying it with Ref2VA as well. Strength 0.7 with 8 steps seems to give decent visual quality, although I can't really comment on the audio yet. Strength 1.0 tends to burn out the image, while 3 steps are basically useless for Ref2VA.
You can try 6 steps regular denoise and then 2 additional steps with this lora. This gave me a results I'm quite happy with.
This is great. I had some trouble with output quality at the start. Took me a while to pin it down to the LoRA loader I was using. https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
Fucking nice bro !!! I use the lora motion booster to help with fast mouvement and its works perfectly. Your the goat ! Hope we could have ref2 version too !!
Which Lora are you referring to as "motion booster"?
Use Tao-Mate for upscaling—3 steps with a denoise setting of around 0.5—and you won't regret it.
yes! +1