AI Photos of You Look Wrong — Researchers Found a Free Fix

AI image generators like Stable Diffusion can be taught to recognize a specific person, pet, or product from just a few pictures — a trick the industry calls "personalization." It's how people make AI avatars of themselves or mockups of their own products. But here's the twist: increasingly, the pictures fed into that training aren't photos from a camera at all. They're images the AI generated itself. This new paper shows that when you train on AI-made images, the results get worse — colors turn oversaturated and details look weirdly crunchy and sharp.

The researchers wanted to know exactly why, so they ran a clean experiment: they took one base model and personalized it twice — once on real photos of a subject, once on AI-generated images of that same subject. The difference traced back to "classifier-free guidance," which is the internal dial that controls how literally the AI follows your instructions. When trained on synthetic images, that dial gets cranked up too high, especially in the fine details. That's what creates the glowing neon colors and the over-sharpened, artificial look.

Their solution, ReGain, is a small correction applied at the moment the image is generated. Think of it like a volume limiter that turns down only the frequencies that got boosted too much. It requires no retraining, no new data, and — importantly — no real photos of the subject. On Stable Diffusion 1.5, it closed 51 to 64 percent of the quality gap compared to a model trained on genuine photos. It also improved results on newer models like SDXL and SD 3.5, without hurting how well the image matches the text prompt.

The catch: this is a research paper, not something you can download and click today. ReGain closes most of the gap, not all of it — real photos still win. But it points to a future where AI avatars, product images, and pet portraits look right even when the only reference material you have was invented by a machine.