FARI: Robust One-step Inversion for Diffusion Model Watermarking

The researchers propose a one-step inversion framework, FARI, combined with lightweight adversarial LoRA fine-tuning of the denoiser to extract diffusion model watermarks, surpassing 50-step DDIM inversion in watermark verification robustness while greatly shortening inference time. The method exploits the fact that the inversion trajectory has lower curvature than the forward generation path and can be approximated with low NFE, enabling end-to-end adversarial training to directly optimize robustness. Fine-tuning takes about 20 minutes on a single NVIDIA RTX A6000 GPU, and the code and pretrained models have been released.