H3 Relay Hires: open-source node upscales anime video 0.5MP to 1080p ~4x faster
crinklypaper · reddit · 2026-10-28
Developer tazmannner open-sourced ComfyUI-H3-RelayHires (MIT, 8 example workflows), a node that finishes MiniMax H3 anime takes at 1920x1088 from a 0.5MP pass, roughly 4x faster end-to-end than a native render.
How it works (one node, three steps):
- Pass 1 at 0.5MP — H3 decides layout, motion, acting and audio; you approve the take here;
- AnimeSR upscale to 1920 (2s) — a small upscaler trained from scratch, fixing the issue that H3 draws 2px ink lines at every resolution, which makes normal upscalers produce overly thick lines and mushy faces;
- One H3 turbo step at 1920 (DMAD LoRA, from sigma 0.60) — redraws detail only; the take and audio match the preview exactly.
On a 5090, 5s clip at 1920x1088: native 500s; 1MP + latent upscale + refine 240s; Relay Hires 118s; with Fast Tail Sampler 107s. Line shimmer 0.13 (0.10 native, 0.22 for the 1MP route); smeared fast frames and tiny faces come out clean.
Bonus Fast Tail Sampler: H3 decides the take in its first 12 of 20 steps, so those run as euler and the last 8 become 4 second-order steps — pass 1 drops from 58s to 47s with the same take.
Honest limits: ink lines slightly bolder than native; very fast action that melts at 0.5MP only gets motion blur (workflow 2 adds a de-rope fix); 2D anime only. Requires MiniMax H3 in ComfyUI, Kijai's DMAD turbo LoRA and taeh3 loader. Tutorial video co-made with Claude.
More from coding & agent
- Grok Bot gets a free AI phone front desk that reads texts, voicemails and drafts replies — elonmusk · 2026-10-28
- Sourcing agent team splits work 80/20: agents automate, humans keep negotiation and final calls — nordic_ash · 2026-10-28
- Try SPARQL locally in five minutes with Oxigraph's serverless Python triple store — JeremyCMorgan · 2026-10-28
- Build One Folder That Teaches Claude Your Entire Business — aryanXmahajan · 2026-10-28
- 41 leftover git worktrees eat 45GB: fix it with a cleanup rule in AGENTS.md — dotey · 2026-10-28
- One prompt fixes AI-generated duplicate code: ask it to dedupe and simplify the architecture — facontidavide · 2026-10-28