NVIDIA open-sources 4K video regeneration for MiniMax H3

The loudest news about a Chinese video model this week came out of an American lab. That is the compliment MiniMax H3 has been waiting for.

The Efficient-Large-Model team — NVIDIA and MIT researchers — pushed weights for Open-H3-Regenerate-2K to Hugging Face, with inference code in the NVlabs/Sana repo. The pitch sounds boring: take a rough clip, rebuild it at high resolution. In practice this is exactly the stage MiniMax keeps locked behind an API.

The full H3 system has three parts: multimodal context handling, base audio-video generation, and high-resolution regeneration. The open H3-Base weights give you 768p with sound. MiniMax's own H3-Regenerate-2K module is API-only. The new project plugs that hole: 2560×1440 out of the box, up to 3840×2160 — native 4K.

What actually got opened

The trick is that almost nothing is retrained. The H3 transformer backbone and the video VAE keep their original weights, frozen. Only a new refinement LoRA gets trained — and that LoRA is what shipped. At inference it merges with community adapters LightX2V FL2VA Turbo and LMS Sharp, then three denoising steps at the target resolution.

Scene, layout and motion come from the draft. Detail is drawn from scratch. That is why the team says regeneration, not upscaling: upscaling inflates pixels, regeneration invents texture.

Speed came from brute force. A community-built latent upsampler blows up the latent representation directly, killing the repeated decode-encode pass at high resolution. Without that shortcut, 4K would crawl on ordinary hardware.

Why this matters more than another model

Anyone can now assemble a bad-clip-to-4K pipeline locally. No API keys, no per-minute billing, no queue behind someone else's inference cluster. For video production that changes the budget line, not the color filter.

Now the position. NVIDIA does not care which model wins. It cares that every pipeline lands on its stack. Taking a Chinese open model, writing one adapter and shipping it in your own repo is a cheap way to plant a flag. MiniMax, meanwhile, loses exclusivity on the most expensive stage of its own system.

The irony writes itself: the loudest coverage of a Chinese model came from a US lab. While everyone argues about who leads, someone else takes the open weights and turns them into a product.