Open-H3-Regenerate-2K: Generate 22-second 2K video on a single B200

An open-source pipeline that replaces the closed-source H3-Regenerate-2K API based on MiniMax-H3. It takes a 768P (1344×768, 121 frames) draft video as input and performs a 3-stage refinement process to regenerate it at 2560×1440 resolution. It is compatible with the Diffusers library and can be run directly in a local environment.

It focuses on performance optimization. Using standard dense attention and BF16 takes 87.26 seconds to process a 5-second clip, but applying Sol-Attn, MXFP8, and CUDA graphs reduces this to 37.39 seconds. Combining this with a high-speed VAE decoder distilled from StreamRender-H3 reduces the total processing time to 22.45 seconds. This is 3.89 times faster than the standard setup.

Three adapters are integrated into a single LoRA file. The Open-H3 refiner, LightX2V FL2VA Turbo, and LMS Sharp are merged along the rank axis to improve inference efficiency while maintaining accuracy. It also includes a 3D-convolution-based latent variable upscaler and a 1.62B parameter non-causal VAE decoder, specializing in high-resolution video generation.

It is suitable for developers with a single B200 GPU who want to reduce high-resolution video generation costs. It can be used in situations where you need to control 2K video generation locally to move away from closed-source API dependencies, or when building workflows to extend existing H3 models to high resolution. It follows the MiniMax H3 Community License, and since licenses differ for each component, verification is required before distribution.