Running NVIDIA's Open-H3-Regenerate-2K on a 4090: sharper but rewrites the actor's performance
friedlc · reddit · 2026-10-29
NVlabs just released Open-H3-Regenerate-2K, an open 2K refiner for MiniMax-H3 drafts whose official code needs 150 GB of VRAM. The author used Claude Code to port its LoRA into ComfyUI and benchmarked it on a single RTX 4090 against SeedVR2.
Findings:
- Default strength (σ 0.909) is very crisp but regenerates the person: younger, eyes open and talking, tie gone.
- σ 0.8 restores the yawn and tie but the tired eyes "wake up".
- σ 0.7 is close to the original but softer than SeedVR2.
- SeedVR2 remains the sharpest option that stays faithful to the source performance.
Upsides: steadier than SeedVR2 (less texture boiling), no color shift, and 2–4 min per 5 s clip on a 4090. Great default for landscapes/animals/products; for acting-driven close-ups the author sticks with SeedVR2.
ComfyUI recipe: H3 FL2VA checkpoint plus LBH latent upscaler; converted LoRA at 1.0, Euler, CFG 1, ManualSigmas 0.909375, 0.857143, 0.666667, 0 (gentler: 0.8, 0.685714, 0.48, 0); Diffusers-format LoRA needs conversion (q/k/v merged into fused qkvproj, MLP fc1 halves swapped); the extra 1.2 GB "Sharp curve branches" file is all zeros and can be skipped.
More from coding & agent
- anymove: motion capture for any creature from one AI video, $0.26 per move — majidmanzarpour · 2026-10-29
- Claude Code plugin makes complete music videos: beat analysis, footage, GPU compositing — AIandDesign · 2026-10-29
- How I AI agent bench: Grok Bot wins big at 93%, Claude Duet trails at 79% — lennysan · 2026-10-29
- RelPrim 1.0 open-sourced to tame retries, rate limits and circuit breakers for LLM APIs — bartrozycki · 2026-10-29
- Claude Code audit uncovers 14,219 secret front-camera photos hidden in Android cache — natesiggard · 2026-10-29
- Why parallel Neo4j imports run slower than serial: lock contention and the fix — JeremyCMorgan · 2026-10-29