Open-H3-Regenerate-2K Matches Proprietary Video Upscaler at 5× the Speed

Scroll SoL-Engine has released Open-H3-Regenerate-2K, an open-source video refinement model that takes low-resolution video drafts — from 384P up to 768P — and regenerates them at full 2K or 4K resolution in just three diffusion steps. The model keeps the H3 transformer and video VAE frozen, adding a single refiner LoRA trained on real high-resolution clips paired with degraded versions. A community 3D-convolutional latent upscaler (H3-Latent-Upscaler) enlarges the encoded draft directly onto the target grid without a decode/re-encode round-trip, and three Euler denoising steps then redraw surface detail at full resolution.

On RefinerBench-150, Open-H3-Regenerate-2K outscores the proprietary MiniMax H3-Regenerate-2K API on all three UniPercept axes as well as VBench aesthetic quality, imaging quality, and subject consistency — while running 5× faster on a single NVIDIA B200 GPU (22.45 s vs. 87.26 s for a 5-second 2K clip). In a blinded user study, the open model was preferred in 87% of 200 judgments. The release directly addresses a gap in the H3 open-source ecosystem: the official API requires unmodified H3 768P input and returns only 2K, while this open alternative accepts drafts from H3 or other video models and targets 768P through 4K.