ReGain: training-free fix restores subject fidelity lost from personalizing on synthetic images
UIUC-CS · hf · 2026-10-09
UIUC researchers show that DreamBooth personalization on diffusion-generated synthetic images degrades subject fidelity with oversaturated colors and excess high-frequency detail. They trace the cause to classifier-free guidance: the angle between conditional and unconditional noise predictions inflates toward high frequencies.
ReGain is a training-free, sampling-time correction that measures how much each frequency band of guidance is inflated relative to the base model and scales it down, requiring no real photos. On SD v1.5 it closes 51-64% of the subject-fidelity gap (DINO, DINOv2, CLIP-I), also improves SDXL and SD 3.5, and preserves text alignment on all three backbones.
More from Research
- Cultural Competence in AI paper presented at COLM 2026 — MaartenSap · 2026-10-09
- Arc Institute's PIE model predicts unseen perturbations in virtual cells, 2.7x AUPRC gain, code written by AI agents — surmenok · 2026-10-09
- Snorkel AI Adds Bonus Livestream on How Benchmarks Are (and Aren't) Used in Model Development — dlwh · 2026-10-09
- Odysseys Benchmark for Long-Horizon Computer-Use Agents to Debut at COLM 2026 — rsalakhu · 2026-10-09
- Humans Spot AI-Rewritten Text 64% of the Time in 1,636-Answer Test — dejanseo · 2026-10-09
- MIT TR roundtable: Stanford PhD who used generative AI to design bacteria-killing viruses — nordicinst · 2026-10-09