FAST-DIPS: Adjoint-Free Analytic Steps and Hard-Constrained Likelihood Correction for Diffusion-Prior Inverse Problems

Researchers propose FAST-DIPS, a new training-free diffusion prior method for solving inverse problems that achieves up to 19.5x speedup while maintaining competitive image quality metrics. The method replaces computationally expensive inner optimization loops with closed-form projections and analytic step sizes, significantly reducing the number of required denoiser evaluations.

• →FAST-DIPS eliminates the need for repeated derivatives and inner optimization loops that slow down existing diffusion-based inverse problem solvers.
• →The method uses hard measurement-space feasibility constraints with closed-form projections instead of iterative approaches.
• →Achieves up to 19.5x computational speedup while maintaining competitive PSNR, SSIM, and LPIPS image quality metrics.
• →Includes theoretical guarantees with proven local model optimality and descent properties under backtracking.
• →Supports both pixel-space and latent-space variants with hybrid scheduling options.