Solo creator ships 21-min AI anime episode 2, shares pipeline lessons

Lunesia-shikishiki · reddit · 2026-09-17

A solo creator producing AI-anime pilot episodes released episode 2 (21 min) and shared key pipeline changes between episodes on Reddit:

Character consistency: stopped chasing a 100% lock; fixes non-negotiables per character (silhouette, palette, one signature detail) and reuses approved reference stills as the source for every shot instead of regenerating from text, cutting most re-rolling.
Locations: each new location is blocked once as a reference image; every shot is a reframe of that space rather than a fresh generation, fixing most geometry drift (doors/windows moving between shots).
Regeneration: exact prompts and references stay attached to every shot in the storyboard — redoing an episode-1 shot for continuity took 5 minutes instead of an afternoon.
The bible: world rules, what's revealed, and a "not before episode N" column; process, not models, prevented episode 2 from contradicting episode 1.

Still requiring a human: pacing, deciding what not to show, cutting around mistakes, voice direction. Most extra time on episode 2 went there, since feedback on episode 1 was about flat voices, long walking scenes, and excess exposition — not image quality. The animation side remains mostly closed-source; the author is still looking for an open-source video model that holds a character across a full 20-minute episode.

Original post →

Solo creator ships 21-min AI anime episode 2, shares pipeline lessons

More from Multimodal