MiniMax H3's 2K finish is now an open 3-step pass

MiniMax's 2K finishing pass for H3 used to be something you rented by the job. In late October there is an open version you can download, and the thing standing between it and your render is the license rather than the price.

Open-H3-Regenerate-2K is an open replacement for MiniMax's closed H3-Regenerate-2K API. It takes a low-resolution MiniMax H3 draft and redraws it at 2560 by 1440 in three passes, with the weights on Hugging Face and the code in NVIDIA's NVlabs repository. The repo was created on October 26, 2026, according to the Hugging Face API, and last updated on October 28.

The reason people are asking about it is simple. The draft step was already open. The finishing step was the paid one, and now there is a downloadable version of the same job.

What Is Open-H3-Regenerate-2K?

It is a refinement add-on for MiniMax H3, the open model that makes video and sound together. H3 settles content and motion at low resolution, where sampling is cheap, and a second pass redraws the detail at full size so faces, fabric, fur and foliage hold up. MiniMax sells that second pass as the H3-Regenerate-2K API, which has no released weights. Open-H3-Regenerate-2K does the same step with published files.

The model card describes a three-step pass that regenerates a 768P H3 draft (1344 by 768, 121 frames, 24 fps) at 2560 by 1440. The project page goes further and lists native output at 768P, 1080P, 2K (2560 by 1440) and 4K (3840 by 2160). The card documents 2K only, and 4K is a project-page claim: the page says the latent enlarger is applied twice, each time within its trained range. Clips of 5 to 15 seconds are refined in a single pass, and the page shows a 15-second, 360-frame clip regenerated at 2K in one run.

It is not tied to H3 drafts. The project page says it takes a draft at any low resolution, such as 384P, 512P, 576P or 768P, made by H3 or by another video model. Its side-by-side comparisons include drafts prepared from LTX-2 and SANA-Video at 1024 by 576.

How the Three-Step Pass Works

Open-H3-Regenerate-2K keeps the H3 transformer and video VAE frozen and adds a single refiner LoRA. The pipeline has four moves. The H3 VAE encodes the draft once. A community 3D-convolutional network, H3-Latent-Upscaler, enlarges the latent straight onto the target grid with no decode and re-encode. The enlarged latent is partly re-noised to sigma 0.909 and denoised in three Euler steps (0.909 to 0.857 to 0.667 to 0), with classifier-free guidance off, so each step is one transformer call at the target size. The VAE decoder turns the result back into frames.

The speed comes from merging. For few steps, the refiner LoRA is merged with LightX2V's FL2VA Turbo and the community LMS Sharp adapter into one file. The card says the merged delta matches the three separate adapters to a relative error of 2.2e-8. On the fast path, decoding uses the 24-layer non-causal decoder from StreamRender-H3, which the card says decodes a 2560 by 1440 clip in 1.6 seconds instead of 15.

What You Need to Run It

The weights are published: one merged LoRA, the Sharp curve branches, an author 3D-conv latent upsampler, the fast VAE decoder, and an unmodified copy of the MiniMax H3 Diffusers components. The code lives in the sol-engine branch of NVlabs/Sana, a research repository rather than a ComfyUI node. The code README lists Linux, Python 3.11 and PyTorch 2.11 with CUDA 13.0.

Every timing in the release is on a single NVIDIA B200 (GB200). The card gives no memory figure, so nobody has shown this pass on a desktop GPU. The card also notes that the timings exclude generating the 768P draft, model loading, text encoding and writing the MP4.

The License Is the Catch

The refiner LoRA, the fast VAE decoder and the bundled base model all follow the MiniMax H3 Community License. That license limits everything to its Applicable Territory, defined as worldwide minus the Excluded Territories, and the Excluded Territories are the European Union, the United Kingdom, the Republic of Korea and the United States of America.

The restriction follows the output. Section V.4 of the license says: "You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory." So for a creator in Los Angeles, London, Berlin or Seoul, this is not a tool to ship with, and the clips it produces are not free to use either. The same section says any such use "is not authorized by this Agreement."

It is also not one license. The card says the bundle is not covered by a single blanket license, because the parts the team did not train follow their own terms: the latent upscaler, and the LightX2V FL2VA Turbo and LMS Sharp LoRAs merged into the main file.

How It Compares to MiniMax's Own API

MiniMax's regeneration endpoint is narrower. Its API reference describes the job as regenerating a source video that meets the MiniMax-H3 768P output specifications into a 2K video, and the project page says the official API "only accepts unmodified H3 768P output." The open version takes a draft at any low resolution and returns any of four sizes up to 4K.

On the authors' own benchmark, the open pass also scores higher on look. Both systems were run on RefinerBench-150, a 150-video set, at 2560 by 1440. On UniPercept, which scores perceived quality out of 100 on three axes, the project page reports averages of 53.4 for the official API and 60.0 for theirs. On VBench it reports higher aesthetic quality, imaging quality and subject consistency, and slightly lower motion smoothness and dynamic degree.

What the Authors Claim, and What Nobody Has Checked

Every performance number in this release is the authors' own, and none of them has been reproduced outside the team. The headline figure is 22.45 seconds for a 5-second 2K clip on one B200, against 87.26 seconds on their dense baseline, which they call 3.9 times faster. That is warm, the median of three runs, and it leaves out the draft, the loading and the file writing. The code README adds that the fast decoder's first call for a new latent shape takes about 39 seconds while it builds the engine and records a CUDA graph.

The quality claims are theirs too. The benchmark scores come from the team's own runs, and the 87 percent figure comes from a blinded user study the team ran: 50 cases, 20 sessions, 200 judgments, with their output preferred in 174 of them. The page lists the authors as Sensen Gao and Haozhe Liu with equal contribution, plus Yitong Li, Haopeng Li, Tian Ye, Jia-Wang Bian, Song Han and Enze Xie.

What Comes Next

The pattern here is the paid layer going open. The draft step was free, the finish was rented, and now the finish is a download, which weakens the case for drafting locally and finishing in the cloud. The two things that decide whether a given creator can use it are unchanged: whether their country is in the permitted territory, and whether they have the hardware the release was timed on. The project page already frames the next step as more of the same, with the SoL-Refiner stages that produce a one-step refiner listed as not used yet.

The weights are on Hugging Face, and the code is in the sol-engine branch of NVlabs/Sana.