TS-LFO: Bypassing Copyright Protection in Diffusion Model Customization via Two-Stage Latent Feature Optimization

Researchers propose a copyright theft attack, TS-LFO, which bypasses copyright protection in diffusion model customization through two-stage latent feature optimization. The method first aligns latent encodings with the input image in the latent denoising stage and suppresses high-frequency noise introduced by defenses, then restores low-frequency semantics with pixel-level constraints in the latent reconstruction stage. Experiments show it can stably bypass SOTA copyright defenses and outperforms copyright attacks such as DiffPure, GrIDPure and IMPRESS under various settings.