A new technique called ReGain fixes a glitch that creeps in when AI image generators are fine-tuned on their own synthetic pictures.
Researchers found that DreamBooth personalization, the standard method for teaching a diffusion model to render a specific subject from a handful of images, degrades badly when those training images are AI-generated rather than photographed. The flaw traces to classifier-free guidance: models trained on synthetic photos show a much wider gap between their conditional and unconditional noise predictions, especially at high frequencies, producing oversaturated colors and extra fake detail. The team built ReGain, a training-free fix applied during sampling that measures how inflated each frequency band of that guidance signal is and scales it back down, no real photos required. Tested on Stable Diffusion v1.5, ReGain closed 51 to 64 percent of the fidelity gap versus models trained on real photos, with similar gains on SDXL and SD 3.5.
As more training data for personalization tools comes from generators rather than cameras, this kind of self-referential drift becomes a structural risk, not a one-off bug. ReGain is notable because it needs no real reference photos at all, and it runs only at sampling time, meaning it could be bolted onto existing personalization pipelines without retraining anything.
It is also a tidy demonstration of a broader problem facing generative AI: models trained on their own outputs tend to drift, and catching that drift early will matter more as synthetic images flood the training pool.