Open-H3-Regenerate-2K regenerates videos in 4K with only three diffusion steps

Open-H3-Regenerate-2K regenerates videos up to 4K with three diffusion steps, downloadable weights, and accelerated 2K processing.

Turn a low-resolution video into a 4K sequence whose details are regenerated by AI rather than simply enlarged. Researchers working on NVIDIA's video generation technology have introduced Open-H3-Regenerate-2K, a system that reconstructs video detail using just three diffusion steps. The model supports output at up to 3840 × 2160 pixels, processing clips between five and fifteen seconds long in a single pass.

The project offers downloadable weights as an alternative to MiniMax's proprietary H3-Regenerate-2K API, which enhances videos generated by the H3 model. Unlike the official API, which accepts specific H3 outputs and produces 2K video, the new system can also process footage generated by other video models.

The researchers also report significant inference acceleration. On a single NVIDIA B200 GPU, processing a five-second video at 2K resolution takes 22.45 seconds instead of 87.26 seconds with SoL-Engine optimizations and a specialized video decoder. Reconstructing detail instead of simply resizing frames

Video generation models often produce their initial sequences at relatively modest resolutions. At this stage, the model establishes the composition, characters, movement, and camera trajectory.

Higher-resolution detail is added later.

Conventional upscaling increases the number of pixels, but it cannot necessarily reconstruct textures missing from the original footage. Hair, fabric, faces, and foliage may appear soft or artificial.

Open-H3-Regenerate-2K takes a different approach.

The system retains the broad structure of the source video, then processes its internal representation through another generative refinement stage. The model attempts to reconstruct visual detail appropriate for the target resolution without creating a completely different scene.

This is not simply sharpening or traditional resizing. It is generative video refinement conditioned on the original footage.

Maintaining movement and character identity remains essential. A system that produces visually convincing individual frames can still introduce inconsistencies between consecutive frames. From 384p to 4K, including footage from other models

MiniMax's official H3-Regenerate-2K API accepts native 768p H3 videos and returns a 2K result.

Open-H3-Regenerate-2K supports a broader range of inputs and outputs.

It can process lower-resolution footage at 384p, 512p, 576p, or 768p and regenerate it at several target resolutions.

Output resolution · Dimensions 768p · 1344 × 768 1080p · 1920 × 1080 2K · 2560 × 1440 4K · 3840 × 2160

The model supports clips from five to fifteen seconds in a single refinement pass.

The team demonstrates a fifteen-second video containing 360 frames, regenerated at 2K without splitting the sequence into independent segments.

Examples also include footage originating from LTX-2 and SANA-Video, rather than MiniMax H3 alone.

For studios using multiple generative video platforms, this compatibility could make it possible to introduce a common finishing stage across different production pipelines.

However, compatibility with multiple sources does not guarantee identical quality. Input stability, compression artifacts, motion complexity, and the underlying model's visual characteristics can all affect the result. Three diffusion steps to regenerate a video

The pipeline consists of four main stages: encoding, latent upscaling, generative refinement, and final decoding.

MiniMax H3 uses a video variational autoencoder (VAE) to convert image sequences into a compressed representation known as a latent.

Instead of decoding that latent into frames, resizing them, and encoding them again at the higher resolution, Open-H3-Regenerate-2K uses H3-Latent-Upscaler, a three-dimensional convolutional network that directly expands the latent representation.

This avoids an expensive intermediate operation, particularly at 2K and 4K.

The enlarged latent is then partially re-noised before undergoing three denoising steps.

The model regenerates textures and fine details directly at the target resolution.

The researchers retain the original H3 diffusion transformer, keeping its base weights frozen during the refiner's training.

Only a dedicated LoRA adapter is trained to reconstruct the missing detail.

This approach avoids retraining the entire video generation model. Training on deliberately degraded footage

The researchers train the refiner using high-quality video clips as reference targets.

Each training video is deliberately degraded by reducing its resolution by a factor between eight and sixty-four, then enlarging it again.

The degraded version retains much of the original scene layout and motion, but loses fine visual information.

The refiner learns to reconstruct a representation close to the original high-resolution recording.

Training footage includes resolutions of 1280 × 704 and 2560 × 1440.

The method builds on the first stage of SoL-Refiner, which focuses on high-resolution refinement training.

Later stages of that research, including reinforcement learning post-training and distillation toward a one-step refiner, are not yet used in Open-H3-Regenerate-2K.

For inference acceleration, the researchers combine their refiner with two community-developed components: FL2VA Turbo from LightX2V and LMS Sharp, an adapter designed to improve detail.

Together, these components allow the refinement process to operate with three calls to the main transformer. From 87 seconds to 22 seconds for a five-second clip

The reported performance improvements come from several optimizations integrated into SoL-Engine.

A major component is Sol-Attn, a sparse attention implementation that reduces the amount of computation required for video processing.

The engine also uses MXFP8 computation, fused kernels, and CUDA graphs to reduce repeated execution overhead.

Finally, the standard MiniMax video decoder is replaced with a faster variant from StreamRender-H3, reduced from 36 to 24 layers.

For a five-second video refined at 2K, the researchers report the following