Instructions to use MATLOWAI/MiniMax-H3-Motion-Adapter with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use MATLOWAI/MiniMax-H3-Motion-Adapter with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline from diffusers.utils import load_image, export_to_video # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("MiniMaxAI/MiniMax-H3", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("MATLOWAI/MiniMax-H3-Motion-Adapter") prompt = "A man with short gray hair plays a red electric guitar." input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png") image = pipe(image=input_image, prompt=prompt).frames[0] export_to_video(output, "output.mp4") - Inference
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
LoRA explained
I had no idea what this lora did.
For those that are confused, I asked GROK to simplify the technical explanation. Hope it helps others.
[ In simple terms: ]
This is a small, specialized “helper” add-on (a rank-16 LoRA) for the MiniMax-H3 AI video model. It is not a general motion booster that makes everything move more.
[ The problem it fixes ]
When people use a special ComfyUI tool called MAINodes Motion Lab, it stretches the timeline of fast action scenes (the “de-rope” pass) so the model has more time to work on quick movements like punches, sword swings, or fights.
Without this LoRA, the model often messes that up:
Motion looks jumpy or “advance-and-snap” (character moves forward, then freezes, then jumps again).
It invents too much extra movement (about 40% more than intended).
[ What the LoRA does ]
It teaches the model to use that extra “stretched time” for smooth, continuous motion instead of inventing extra movement or stuttering.
Result: fast action looks more natural and fluid, closer to the speed the original shot intended.
Important limitations (from the creators)
It only works with the MAINodes de-rope / Motion Lab pass. Loading it on a normal text-to-video generation does almost nothing useful.
Best for fast motion (especially anime-style action). On calm or slow scenes it can over-smooth things or mute colors/particles a bit.
It’s a pilot release (trained on a small set of clips). There’s also a newer “temporal expansion” version trained on real high-speed camera footage.
Bottom line: It’s a targeted fix that makes fast-action regeneration in one specific ComfyUI pipeline look smoother and more realistic, rather than jittery or exaggerated.
You got it! It's particularly important for the windowed mode in de-rope which actually causes weird black flickering without this adapter. It's a part of a series of experiments where I'm generally working towards training the model to be time aware during generation and change its time compression on the fly but I'm thinking it is outside my personal GPU reach... DyRoPE is another leg of the same research and worth looking at if you haven't yet. It's not perfect but it can be handy if you get some strange effects of small objects like birds deciding to flap their wings more quickly based on the hold map. I've found 2 close to optimal settings for it but it still needs work (never enough time as a hobby research).