Open-H3-Regenerate-2K

Open-H3-Regenerate-2K is an open replacement for the closed, hosted H3-Regenerate-2K API. It keeps the H3 transformer and video VAE frozen and adds a refiner LoRA, trained on real high-resolution clips paired with degraded copies. The LoRA is merged with LightX2V's FL2VA Turbo and the LMS Sharp adapter into a single LoRA. It is published on NVIDIA's SoL-Engine research blog (NVlabs Sana project).

At inference the draft is encoded once by the H3 VAE. A community 3D-convolutional H3-Latent-Upscaler enlarges the latent directly onto the target grid. The latent is then re-noised to σ = 0.909 and denoised in three Euler steps with classifier-free guidance off, and the VAE decodes the result. Unlike the official API, which accepts only unmodified H3 768P output, it takes drafts at any low resolution (384P to 768P) from H3 or other models such as LTX-2 and SANA-Video. It outputs 768P, 1080P, 2K (2560×1440) or 4K (3840×2160) for clips of 5 to 15 seconds.

The authors report that with SoL-Engine (sparse attention, MXFP8 linear layers, CUDA graphs) and the fast StreamRender-H3 VAE decoder, a 5-second 2K clip takes 22.45 s on one B200, versus 87.26 s without. On RefinerBench-150 they report higher scores than the official API on UniPercept axes and several VBench metrics, and an 87% preference across 200 blinded judgments. These are the authors' own results. One YouTube reviewer noted that the regeneration can change details such as faces, since the diffusion passes redraw surfaces rather than only upscale.