MiniMax H3's 2K Regeneration Stage Is Now Open-Source and 3.9x Faster

Which stage this fills

The full MiniMax H3 pipeline covers multimodal context processing, base audio-video generation, and high-resolution regeneration. The open-weight H3-Base handles 768P audio-video, while the high-resolution H3-Regenerate-2K stage was previously available only through MiniMax's API. Open-H3-Regenerate-2K, from the Efficient-Large-Model team of NVIDIA and MIT researchers, publishes weights on Hugging Face with inference code in the NVlabs/Sana repository — a self-deployable open implementation of that second stage.

How three steps produce the frame

During training the H3 Transformer backbone and the video VAE are kept frozen; only a new refinement LoRA is trained. At inference that LoRA is merged with community adapters such as LightX2V FL2VA Turbo and LMS Sharp, and detail is redrawn at the target resolution in three denoising steps, with the draft's scene, layout and motion information as the basis.

The speed comes from three things: a community-built latent upscaler that enlarges the latent representation directly, avoiding repeated decode and re-encode at high resolution; plus SoL-Engine sparse attention, low-precision computation, and a fast VAE decoder.

Numbers, as published by the team

The team's tests ran on a single NVIDIA B200 after warmup on a roughly 5-second clip: latent upscaling, three-step refinement and 2K decoding took 22.45 seconds in total, against 87.26 seconds for an unoptimized implementation — about 3.9 times faster. The fast decoder cut decoding from roughly 15 seconds to about 1.6 seconds.