Instructions to use MiniMaxAI/MiniMax-H3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MiniMax H3
How to use MiniMaxAI/MiniMax-H3 with MiniMax H3:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Diffusers
How to use MiniMaxAI/MiniMax-H3 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("MiniMaxAI/MiniMax-H3", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Inference
- Notebooks
- Google Colab
- Kaggle
[FL2VA] Anime style videos have half the FPS
Anime style videos generated by MiniMax-H3 are likely to be 12 FPS instead of 24 (the resulting 24 FPS video will have each frame repeated twice), or they can be of variable framerate (where one part is smooth and the other part is stuttering). I can only get like 1 out of 10+ videos that are actually smooth 24 FPS.
I know this is mostly contributed by the training data where most anime is just like this. But I struggle to generate one in real 24 FPS, even by using prompts like "Smooth motion", "Fluid", "High framerate".
Does anyone have any success in generating real 24 FPS anime style videos consistently?
Maybe you could try generating a realistic style video first, then transfering it into anime style.
In fact, even realistic video generated by H3, there are frame extraction and almost identical 2 frames problems. I guess most training data is not native 24fps video but resampled from 30/60fps, which makes the generated video not smooth.
Maybe you could try generating a realistic style video first, then transfering it into anime style.
That's one way to go about it, but we could never guarantee the second style transfer pass to not also generate at 12FPS.
I'm thinking to just select every 2 frame and slap GIMM VFI on it, that might be easier as a workaround. Though I still hope this can be fixed at prompt / model level.