Timestep Rescheduling in Diffusion Inversion
Abstract Overview
This paper studies how the choice of diffusion timesteps affects inversion fidelity in deterministic diffusion inversion. The authors show that inversion error depends strongly on timestep size and follows a parabolic pattern across the trajectory, with larger deviations at both early and late stages. Based on this analysis, they propose TRDI, a nonuniform timestep rescheduling method that combines global timestep rescaling with local dynamic-programming-based refinement. The method is designed as an off-the-shelf scheduler change for existing inversion pipelines and is reported to require no extra parameters and essentially no online runtime overhead because the dynamic programming schedule can be computed offline.
Novelty
The paper's main novelty is to treat timestep selection itself as a first-class factor in diffusion inversion rather than only refining per-step inversion solvers. It derives a timestep-dependent inversion error formulation linked to a scaled fixed-point problem and uses that analysis to build a coarse-to-fine rescheduling strategy with global rescaling and local dynamic programming.
Results
Across image reconstruction on MSCOCO and image editing on PIE-Bench, the proposed scheduler consistently improves multiple inversion baselines including DDIM, ReNoise, NPI, and GNRI. On MSCOCO with SD v1.5, for example, adding the scheduler improves PSNR, SSIM, and LPIPS for all four baselines, while editing experiments on SDXL and SDXL Turbo show better structure distance and background preservation with largely comparable CLIP similarity. The dynamic-programming stage is reported as negligible in practice, taking 9.32 seconds offline for a 50-timestep schedule versus nearly 3 hours for the full reconstruction dataset.
Key Points
- The paper identifies a timestep-dependent inversion error pattern, with larger errors for large steps and a parabolic dependence on timestep position.
- TRDI reallocates a fixed inference budget through global timestep stretching and local dynamic-programming rescheduling, without changing the underlying inversion method.
- Experiments show consistent gains in reconstruction and editing quality across several inversion baselines and diffusion models, especially for preserving structure and background details.