If you make images with AI, you know the specific pain: the first generation is easy. Editing it is where everything falls apart.
I make YouTube thumbnails and playlist covers. I'd get a result I liked, ask to change one small thing — "make the jacket blue" — and get back a completely different picture. New face. New background. New lighting. The one thing I asked for, plus ten I didn't.
After the hundredth time, I realized the problem isn't the model — it's the prompt. When you say "change the jacket," the model has no idea what you want to keep, so it re-rolls the whole scene. The fix is to spell out, every time, what stays locked and what changes, in language the model actually follows.
That's tedious by hand, so I turned it into a plain-text bundle you paste into ChatGPT or Claude. You describe what you want; it writes the direction — one ready-to-paste prompt that says exactly what to preserve and what to change. No code, no API, nothing to install.
Clearest example — same street photo, one instruction ("add falling snow"):
The snow appears; the houses, the road, the parked vans stay exactly as they were. Most tools hand you a brand-new street.
It can also diagnose a render that came out wrong — what's off, why, what to change, and what to keep:
How it works (no magic): it's plain-text reasoning, not a model or an integration. It works out how much detail a request needs, separates what's immutable from what you're changing, and phrases the result for the renderer (ChatGPT / GPT Image by default). Because ChatGPT re-renders the whole frame instead of inpainting, the trick is to restate the locked parts explicitly.
Honest limits: it's tuned and tested on ChatGPT / GPT Image; the Midjourney, Gemini, Ideogram and FLUX profiles are starting points, not verified. And no prompt makes a whole-frame renderer do true local edits — it's a strong mitigation, not a guarantee.
The free version is on GitHub and takes about a minute to try. There's a $19 one-time Pro bundle with the full preservation system, the edit-state ledger, precise single-element edits and deep diagnosis — but the free one shows the whole idea.
Free on GitHub: https://github.com/leewilliam200/visual-director-lite
Would genuinely love feedback — especially where the "keep everything else the same" logic breaks for you. What's the one thing you always want an AI editor to preserve that it never does?
Top comments (0)