A few prompts in, the face was wrong. Not dramatically. The jaw sat differently. The eyes had drifted a shade. Panel one, I’d have caught it anywhere. Panel four, a stranger wouldn’t have noticed.
Gemini did it. Then ChatGPT. Then Grok. Then Meta. Four companies, four training runs, one identical failure. That’s not a prompting problem. That’s a wall.
I had a canon doc by then. Continuity rules, story arcs, voice — a whole Notion OS treating the comic like a product. None of it ever went into an image prompt. Not once. I’d built infrastructure for the half of the problem I understood, and left the half that actually broke the thing untouched.
Text remembers. Faces don’t. Not yet.
The blog comes first. Voice holds in a paragraph before it holds in a panel — and a paragraph doesn’t need a canon doc to keep it steady. Once enough posts prove the voice holds on its own, I go back and hand the system the piece I skipped the first time.
Leave a Reply