FARI: Robust One-Step Inversion for Watermarking in Diffusion Models

Researchers introduced FARI, a fast asymmetric robust inversion method for diffusion-model watermarking.

It uses one-step inversion plus lightweight adversarial LoRA fine-tuning to extract watermarks more efficiently.

The paper says FARI outperforms 50-step DDIM inversion on verification robustness after about 20 minutes of fine-tuning on a single GPU.

The approach could make watermark authentication faster and more practical, while enabling robust adversarial training.