An unofficial ComfyUI port of the FastH3 4-Step Preview LoRA.
Sources
Original LoRA (dense-datafree): FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA
Conversion recipe: taken from the metadata of lightx2v/Minimax-h3-Turbo
Conversion recipe (as recorded in the source metadata):
format = pt
source_format = Diffusers PEFT LoRA
target_format = ComfyUI generic LoRA
qkv_fusion = block diagonal B; concat A; alpha multiplied by 3
training_rank = 128
training_scale = 1.0
training_alpha = 128.0
swi_glu_mapping = Diffusers [value;gate] -> ComfyUI [gate;value]
base_model = Comfy-Org/MiniMax-H3 minimax_h3_fl2va_bf16.safetensorsCompatibility
✅ I2V and T2V — working on the MiniMax H3 FL2VA model (I2V 1- and T2V 2-sample videos)
⚠️ Ref2VA — partially working, but loses significant prompt adherence; an official release is recommended
Examples are made on default minimax workflow, 4 steps, Larryvrh turbo sampler
8ish steps making sound better in my own experience.
Workflows are imbedded in the videos.
Credits
All credit for the original model and the conversion recipe belongs to the original authors:
FastVideo: https://github.com/hao-ai-lab/FastVideo
MiniMax-H3-Turbo: https://github.com/ModelTC/Minimax-H3-Turbo#model-specs
This conversion is an independent, unofficial effort and is not affiliated with, endorsed by, or sponsored by FastVideo, lightx2v, or ModelTC.
Right now I know about an experimental checkpoint by Kijai, would love to hear the difference from users.
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main
Description
FAQ
Comments (17)
There's only 1 video and the WF is not embedded in it
There is, just drag and drop it in comfy, civitai can't read output from Video Combine node.
Second example in moderation for some reason, but it's the same workflow just without the first_frame
Oh my god, this is really amazing, thank you so much for such a great turbo!
Thanks. Not mine though, leave a star on https://github.com/hao-ai-lab/FastVideo
4070TI 12GB 32G,0.8MP,4STEP,10S VIDEO,250S-300S
0.8MP,8STEP,10S, 500S
thanks for benchmarking
@NoiseWriggler Tested on the hybrid_fl2va_ref2va_b25-49 INT model; it works just as well.
@FourBunny oh that's very cool, thank you
A quick question: is this LoRA meant to be used with the "minimax h3 fast video vsa" model or the "minimax h3 standard" model?
with the original one, I've personally tested on "minimax_h3_fl2va_pruned_fp8_scaled"
@NoiseWriggler Ok! Thanks for the quick reply.
sampler and scheduler? thx
https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
simple scheduler
Just a quick heads-up for everyone—but a very important one.
This LoRa completely changes the prompt understanding and the aesthetic of the H3 model.
I’ve just run about 30 tests using various turbo LoRas, fixed seeds, 0.4 and 0.5 MP settings, different prompts, 10-second clips, T2V, and R2V.
So, if you’re generating content with this LoRa simply because the results look good—without having done A/B testing against other turbo LoRas—
you’re missing out on completely different, much more beautiful videos and better prompt adherence.
Bingo. A lot of loras here, especially those trained using SDXL assets, are ruining the model and people complain about it sucking. And for those loras that aren't using SDXL assets, people run them at full power. They don't understad that they need to lower or raise weights.
LoRAs will affect the model, that's the point, but you don't want them to change things so much that they ruin the original model's quality. If the lora says it's a "realistic" lora and it only makes every characters skin look smooth and plastic like... then it's not realistic. People don't have smooth skin. Period. Get off the internet, stop looking at instagram filtered faces, and look at real people.
And if the lora has a bunch of samples that all have same face syndrome? AVOID.
of course,it hightly depend on the prompt, like with all model the prompt is the most important base for a good generation.
But based on my tests, when you prompt need a particular aesthetic, vivid colors, etc...the difference between this turbo lora and other is big.
It will not make significant difference for simple prompt, simple scene.
This lora ruining image, it become plastic and fried.