Minimax-h3_Singularity: Generate HDR-quality action videos with 4-step inference

This fine-tuned model is based on MiniMax-H3 and synthesizes videos in the ComfyUI environment from various inputs such as text, images, and reference videos. It underwent three days of weight optimization to remove artifacts caused by high-step training while maintaining the versatility of the original model.

Trained on an HDR dataset, it reduces motion blur during fast movements and corrects facial distortion of distant subjects. It eliminates unnatural skin gloss to preserve realistic textures and shows particular strength in complex action sequences such as sword fighting or magic effects, as well as camera work control.

Using a 4-step fine-tuned LoRA alongside the model enables fast generation without quality degradation for higher inference speed. An online demo is also provided for immediate testing in a browser without local GPU setup, and commercial use is permitted under the Apache 2.0 license.