Diffusion Model Inversion

Diffusion model inversion works backward from an image, audio recording, or other observation to estimate the noise or trajectory that a diffusion or flow model can use to reproduce it. Approaches include DDIM and ODE/SDE inversion, fixed-point and gradient-based optimization, and trajectory or timestep corrections; some methods also recover prompts or optimize the starting noise. Inversion supports text-guided editing, image restoration and reconstruction, watermark detection, and scientific inverse problems, with evaluations often measuring reconstruction fidelity, edit alignment, and preservation of unchanged regions on benchmarks such as PIE-Bench. A central tension is that a close reconstruction does not necessarily mean the original latent was recovered accurately, while stronger guidance or edits can reduce fidelity and structural preservation.