Open-H3-Regenerate-2K: open 3-step diffusion refines video drafts up to native 4K, 5x faster than the API
linoy_tsaban · x · 2026-10-28
The SoL-Engine team released Open-H3-Regenerate-2K, an open replacement for MiniMax's closed H3-Regenerate-2K API: a 3-step diffusion pass that regenerates low-resolution video drafts at high resolution, with native output from 768P up to 4K (3840×2160) for 5–15 second clips.
How it works:
- Freezes the H3 transformer and video VAE, adding a single refiner LoRA trained like Stage I of SoL-Refiner on real high-res clips paired with degraded copies;
- At inference the draft is encoded once; a community 3D-conv network, H3-Latent-Upscaler, enlarges the latent directly onto the target grid with no decode/re-encode;
- The upscaled latent is re-noised to σ=0.909 and denoised in three Euler steps (0.909→0.857→0.667→0) with CFG off — each step is one full-resolution transformer call;
- For these few steps, the refiner is merged with LightX2V's FL2VA Turbo and the LMS Sharp adapter into one LoRA.
Performance: the official API only accepts unmodified H3 768P output and returns 2K; this project takes 384P/512P/576P/768P drafts from H3 or other models and runs 5x faster than the official API on a single B200. Further speedups via SoL-Engine (Sol-Attn sparse attention, MXFP8 linears, CUDA graphs) and StreamRender-H3's fast VAE decoder. Code and weights are on GitHub/HF.
More from Multimodal
- Each Labs launches Lockstep, a video recast model with locked character motion — miilesus · 2026-10-28
- How those AI Tiny Desk pet band videos are made: a 4-step casting-and-motion workflow — Hungry_Incident1146 · 2026-10-28
- Back to ComfyUI: best H3 turbo LoRAs and how Open H3 Regenerate-2K upscaling works — truci · 2026-10-28
- AI-generated influencer travel vlog captures a full day from sunrise to dusk — MillionaireMatchapp · 2026-10-28
- Developer turns a few Airbnb photos into a full 3D walkthrough with Claude — CodeByPoonam · 2026-10-28
- MOSS-Image character LoRA trained on RX 7900 XT 20GB in 3.5 hours on native Windows — iBorgSimmer · 2026-10-28