New FastH3-Omni

FastH3-Omni is an eight-step refined version of the MiniMax-H3 video model developed by the FastVideo team. You can directly try it out online via this Hugging Face page.

The process involves uploading images, video clips, and audio files in sequence as reference materials, along with a description of the scene. The system will then combine these references and descriptive tags to generate a short video accompanied by an appropriate soundtrack. Using labels such as “Image 1” or “Video 2” in the descriptions helps ensure that each element is correctly matched with its corresponding media file, making it much easier to manage than using pure text descriptions alone.

In terms of speed, the original version requires dozens of sampling steps, while this version only needs 8 steps. The demonstration uses the FP8 quantized version; the official team has also made available the NVFP4 version and the ComfyUI workflow. Those who have a graphics card at home can download these resources and test them themselves. The online version runs on ZeroGPU – it’s free, but there may be a waiting list. It’s ideal for those who want to see the results first before deciding whether to deploy it locally.

Before using it to shoot vertical short videos, you can…TVLabObserve how popular dramas with similar themes begin, and at which episode the plot takes a turn. Then pick an inspiration card that contains a detailed breakdown of the scenes, and use it to generate each element of the story one by one, corresponding to the order of the shots.