Open-H3-Regenerate-2K: open 3-step video refinement up to 4K, 5x faster than MiniMax API
songhan_mit · x · 2026-11-10
The SoL-Engine team released Open-H3-Regenerate-2K (plus a new IC version) as an open replacement for MiniMax's API-only H3-Regenerate-2K stage.
- Approach: freezes the H3 transformer and video VAE, trains only a plug-and-play refiner LoRA; the IC version keeps the low-res draft in context so faces, lip sync, text, and layout stay faithful
- Pipeline: a community 3D-conv H3-Latent-Upscaler enlarges latents directly to the target grid (no decode/re-encode), then re-noises to σ=0.909 and denoises in 3 Euler steps with CFG off
- Results: 5.0x faster than the official API on a single B200; native 768P–4K output for 5–15s clips, flexible 3/4/8-step inference, and 384P→768P upscaling
Code, weights, and project page are all open.
More from Multimodal
- Fable 5.2 nails 'the video of its dreams' in just two shots — DeryaTR_ · 2026-11-10
- Z-Image Turbo image comparison with prompt and workflow shared — FotografoVirtual · 2026-11-10
- A Single Prompt With Fable 5.2 Produced This Surprisingly Polished Pixel Action Game — MellowShader · 2026-11-10
- Generating game music with Fable 5.2 in just 3 prompts — petergyang · 2026-11-10
- Video-Coder trains a coding agent for video generation, gaining 24+ points on PhyGenBench — HuaxiuYaoML · 2026-11-10
- Open-source Gemini Flash Image MCP Server adds text-to-image and editing to agents — modelcontextprotocol · 2026-11-10