Trust-region Diffusion Models with Entropic Constraints for Zero-shot Inverse Problems

Key Points

• Restoration performance improves with the proposed trust-region diffusion sampler, ensuring stability in zero-shot inverse problems.
• The Kullback-Leibler divergence-based method effectively maintains a high-density region for prior distributions.
• Observational analysis compares the TRIDENT model against existing diffusion methods, revealing significant enhancements in perceptual quality.
• Results support the effectiveness of entropic constraints in achieving reliable convergence and operator invariance.