One-pass 15s anime fight with MiniMax H3 ref2va: what's the next quality lever?
lajonquillebleu · reddit · 2026-10-06
A Reddit user shares a full ComfyUI setup for generating a 15-second 2D anime fight scene in one pass with MiniMax H3 (ref2va) — 362 frames @24fps, 28 min per clip on an RTX PRO 6000 — and asks the community where to push quality next.
What already worked:
- Prefacing detaileddescription with "target video is 2d colored anime" prevents drift to generic 3D — the single biggest win
- Real anime screencaps as references, each assigned an explicit role in the prompt (face / expression / fighting posture / set / framing)
- Cutting rhythm from 4 to 8 shots per 15s (1.8s each) with impact verbs and material consequences made the fight read much better
Open questions: soft motion on some hits, uneven shot-to-shot quality (some go plasticky), whether 8 references is too many, and untapped gains at the current resolution/step count. A valuable first-hand workflow log for multi-shot AI anime.
More from Multimodal
- Rainy convenience-store video demo shows off "GPT-6 ASTRA" generation, prompt included — iamfakhrealam · 2026-10-06
- OmniChar adds consistent voice cloning: 10-second sample locks face, body and voice — ashishsanu · 2026-10-06
- OmniChar .char format adds consistent voice cloning from a 10-second sample in ComfyUI — ashishsanu · 2026-10-06
- MiniMax H3 seamless long-video workflow: dual sampling, color transfer and sharper latent upscaler kill visible seams — wjc_5 · 2026-10-06
- Runway announces World Runner: generative worlds on a pocket Game Boy-style console — c_valenzuelab · 2026-10-06
- Gothic vampire short film made with Seedance 2.5 on Runwayml — azed_ai · 2026-10-06