Instructions to use lightx2v/Minimax-h3-Turbo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use lightx2v/Minimax-h3-Turbo 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("lightx2v/Minimax-h3-Turbo", dtype=torch.bfloat16, device_map="cuda") pipe.to("cuda") prompt = "A man with short gray hair plays a red electric guitar." image = load_image( "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png" ) output = pipe(image=image, prompt=prompt).frames[0] export_to_video(output, "output.mp4") - Notebooks
- Google Colab
- Kaggle
8-step V1.0 test results
We’re working on improving the visual quality of our generated videos, with a particular focus on fine details.
Will next version be 4 steps again?
What Strength, Sampler, and Scheduler you recommend for ComfyUI ? And thank you for wonderful work!
Thank you for your great work! The newly provided 8-step LoRA this time, version 1.0, is much better than 0.1! Looking forward to REF2VA next.
I’m having some issues with the new 8-step version in ComfyUI (together with the Spectrum Node + SageAttn). The quality feels noticeably worse than with the old version, which I was also using with 8 steps.
I’ve also noticed some very strange behavior in both videos I generated. In the first video, there’s a scene where someone is supposed to be sitting. Instead of actually sitting, she moves around on her knees like a zombie, almost as if she were paralyzed (bad motion).
In the second video, someone is holding an object in their hand—or at least they’re supposed to. Instead, the object moves up and down on its own as if controlled by some invisible force.
Unfortunately, I can’t share the videos here, but I never experienced this kind of behavior with the old version. I’m not sure whether Spectrum might be causing it, but it was working great with the old version.
I'm using it at a strength of 1 with res_multistep / simple.
I'm using it at a strength of 1 with res_multistep / simple.
Maybe Spectrum Node causing it, I tried Euler/Simple, Euler/Beta, Res_Multistep/Simple, and Res_Multistep/Beta and getting good results on both 0.5 and 1.0 strength. I'm using SageAttn so its not an issue but yeah Spectrum might be causing it.
Try removing spectrum node and just generate with 8 steps. 8 step lora is better than 4 steps from these examples
Unfortunately, I can’t share the videos here, but I never experienced this kind of behavior with the old version. I’m not sure whether Spectrum might be causing it, but it was working great with the old version.
.