Nvidia drops an open-source tool to boost AI video to 4K

Open H3 Regenerate 2K is an open-source video upscaling tool from Nvidia that redraws low-resolution AI video drafts at crisp 2K or 4K resolution.

State-of-the-art open models like MiniMax H3 typically generate video in two stages: draft cheap motion at low resolution, then sharpen everything in a second pass. While the H3 base model was released, the official 2K refiner remained a proprietary cloud API. Nvidia built this open alternative so creators can run the full high-resolution pipeline on their own GPUs.

How it turns blurry drafts into 4K

Standard upscalers simply stretch pixels, leaving fine textures muddy. Open H3 Regenerate 2K actually redraws the scene using a diffusion pass, filling in realistic fabric folds, fur, and background foliage.

Nvidia kept the base H3 transformer and video VAE frozen, layering on a single refiner LoRA that merges FL2VA Turbo and LMS Sharp adapters.

Instead of decoding and re-encoding frames, a 3D-convolutional network resizes the latent representation directly onto the larger grid. The model adds a controlled burst of noise and runs just three Euler diffusion steps without classifier-free guidance, keeping compute remarkably low.

Paired with Nvidia's SoL-Engine sparse attention and a fast VAE decoder, refining a 5-second 2K clip takes just 22.45 seconds on a single B200 GPU, down from over 87 seconds on unoptimized setups.

Beyond H3: flexible inputs and single-pass 15-second clips

While the official API strictly demands 768P drafts, Nvidia’s open model accepts inputs at 384P, 512P, 576P, or 768P. It also works out of the box with drafts from third-party generators like LTX-2 and SANA-Video.

Outputs scale natively to 768P, 1080P, 2K, or full 4K. It can process clips from 5 to 15 seconds (up to 360 frames) in a single pass without visible splicing seams.

On the RefinerBench-150 benchmark, it beat the official closed API in aesthetic quality and subject consistency. In blinded human evaluations across 200 pairs, raters preferred Nvidia's output 87% of the time.

There is one caveat: because the model hallucinates new details during diffusion, it may subtly reshape facial features rather than preserving the exact likeness from your draft.

All code, model weights, and setup documentation are now live on GitHub.

Why it matters

This release solves one of the biggest headaches in open-source video: the gap between cheap low-res drafting and cinematic high-res finishing. Local creators no longer have to rely on paid APIs to make their AI clips look professional.

Nvidia's approach is clever because it avoids heavy end-to-end retraining, relying instead on a lightweight LoRA and a 3-step diffusion trick that runs surprisingly fast. While the slight shift in facial features means you should be cautious on character close-ups, it is an essential upgrade for landscapes, motion graphics, and background plates.