Serialized manga made end-to-end by AI agents. Every page passes five gates before it ships — and a human approves all of them.
One page = five sequential stages. Any failure loops the panel back to generation — nothing is patched by hand or spliced from old output.
Dialogue, action beats, and camera notes authored as validated JSON — cause → micro-reaction → action in adjacent beats.
School halls, classrooms, night skies — each location is designed once as a locked reference sheet.
Diffusion models (Krea2 / FLUX / Qwen-Image) on a local GPU cluster, persona locked by LoRA + prompt anchoring.
Vision QA checks continuity, expression, pose dynamics, Korean glyph readability. Pass, or regenerate.
Independent full-size page canvases, stitched in order into one printable manga page. No collage shortcuts.
Persona prompt blocks (face, hair, eyes, outfit, props) are fixed verbatim across every panel — consistency is engineered, not lucky.
Every speech bubble's Hangul is OCR-verified at pixel level; illegible text fails the gate and the panel regenerates.
Height, tilt, and distance are specified per panel. Agents must show action — verbs, not poses.
One color rule per scene: limited palette plus a spot color. Atmosphere stays coherent across a serialized arc.
Every stage checkpoints to disk. A crashed render resumes from the last verified panel — zero re-rolls of approved work.
The final gate is a person. Agents propose; only approved pages are published.