The Efficient-Large-Model team, a joint group of Nvidia and MIT researchers, has published model weights for Open-H3-Regenerate-2K on Hugging Face, with inference code in the NVlabs/Sana repository. The project does one specific job: take a low-resolution video draft and regenerate it at high resolution, at 2560x1440 for 2K output and up to 3840x2160 for native 4K. What makes it notable is the base it stands on. The whole thing is built on MiniMax H3, the Chinese open-weight audio-video model released in August.
The structure of MiniMax's own system has three stages: multimodal context processing, base audio-video generation, and high-resolution regeneration. Only the first two are open. H3-Base, whose weights are public, produces 768p audio-video, while the official H3-Regenerate-2K module has stayed behind an API. The Nvidia and MIT release fills exactly that gap, giving teams a high-resolution refinement stage they can deploy on their own hardware.
The training recipe is deliberately cheap. The researchers froze both the H3 Transformer backbone and the video VAE and trained only a new refinement LoRA on top. At inference that LoRA is merged with community adapters, LightX2V's FL2VA Turbo and LMS Sharp among them, and runs three denoising steps at the target resolution, using the scene, layout and motion already encoded in the draft as its guide. For speed, the pipeline applies a community latent upscaler to enlarge the latent representation directly instead of repeatedly decoding and re-encoding at high resolution, then leans on SoL-Engine's sparse attention, low-precision compute and a fast VAE decoder.
The measured gains are large. On a single Nvidia B200, with warm caches, a roughly five-second clip goes through latent upscaling, three-step refinement and 2K decoding in 22.45 seconds, down from 87.26 seconds for the unoptimized implementation. That is about 3.9x faster end to end. The fast decoder alone pulls video decoding from roughly 15 seconds to about 1.6 seconds.
Quality held up better than expected too. On RefinerBench-150, with identical inputs and identical 2K outputs, Open-H3-Regenerate-2K averaged 60.0 on the UniPercept perceptual metric, against 53.4 for the official H3-Regenerate-2K API. On VBench it scored higher on aesthetic quality, imaging quality and subject consistency, while the official API kept a slight edge on motion smoothness and dynamic degree. These are the project team's own evaluations and have not been independently replicated.
The release is part of a much larger wave around H3. MiniMax opened the weights in August, and within a month the model passed 24 million downloads and spawned more than 300 public derivative models, making it one of the most downloaded models of 2026. The Nvidia and MIT team alone has shipped several H3 derivatives in different directions: StreamRender-H3, a streaming video rendering model built on H3 Ref2VA, and Sol-H3-VAE, a fast decoder distilled from the H3 video VAE.
The surrounding ecosystem has filled in just as fast. LightX2V shipped four-step and eight-step distilled versions under the MiniMax-H3 Turbo label. A community fine-tune called Singularity concentrated on motion, facial detail and visual style. FastVideo and Pika released FastH3 and the six-step Pika H3 Flex for international users, while ComfyUI and Unsloth adapted weights and quantized builds to bring H3 into local creative workflows.
The strategic reading is straightforward. MiniMax built a strong high-resolution stage and kept it as a paid API; an American lab has now reproduced an open substitute on top of the Chinese base model, and by its own measurements made it slightly better on perceptual quality while running several times faster than the naive path. That is the open-weight flywheel working as advertised, and it puts pressure on every vendor whose business model depends on holding back the last mile of a pipeline. For production teams the practical question is no longer whether high-resolution regeneration can be self-hosted, but how quickly the self-hosted version catches up each time the base model changes.
Comments (0)
Log in to join the discussion
Log InNo comments yet