ReGain: Restoring Subject Fidelity in Personalization on Synthetic Images
Fine-tuning DreamBooth on synthetic subject images can distort the subject, and ReGain corrects part of that loss at sampling time.
The paper traces the fidelity drop to inflated classifier-free guidance, especially in high-frequency bands. ReGain compares that guidance inflation against the base model and scales the affected frequency bands down without needing real photos. On Stable Diffusion v1.5, it closes 51-64% of the measured subject-fidelity gap, and the authors report gains on SDXL and SD 3.5 while preserving text alignment. HF Daily Papers' note
The paper traces the fidelity drop to inflated classifier-free guidance, especially in high-frequency bands. ReGain compares that guidance inflation against the base model and scales the affected frequency bands down without needing real photos. On Stable Diffusion v1.5, it closes 51-64% of the measured subject-fidelity gap, and the authors report gains on SDXL and SD 3.5 while preserving text alignment. HF Daily Papers' note
score 4