Teaching one image AI to edit and generate without the skills fighting each other

Teach an AI to edit photos, and it often gets worse at making them from scratch. Teach it to do small local edits — change a hat's color — and it stumbles on big global ones like turning day into night. These skills don't naturally coexist in one model. They compete.

DanceOPD is a training method that lets them share. Here's the idea. Picture each skill as a separate "expert" — a text-to-image expert, an editing expert, a realism expert — each one essentially a guide that, at any half-finished stage of an image, points which direction to push the pixels next. (Image generators build pictures by repeatedly nudging random noise toward a clean result; each expert knows which nudge to make.)

The trick is on-policy learning. Instead of drilling the student model on the experts' own example images, DanceOPD lets the student produce its own half-finished images, then asks the relevant expert: from exactly where you are now, which way should you go? The student copies that direction. Because it learns on the states it actually visits — not idealized ones — the skills stop interfering.

One method routes each training sample to the right expert and matches directions with a single simple error measurement. It even absorbs a standard sharpening trick used during generation directly into the model.

The paper reports better skill-combining while keeping core quality, but offers no head-to-head numbers here, so how big the gain is stays unclear.

Why you should care: The image tool that edits your photo and the one that generates it from a prompt could finally be the same model — without one skill quietly sabotaging the other.