arxiv:2605.26470

Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems: Optimizing Guidance and Stochasticity Schedules

Published on May 26

Upvote

Authors:

Junseo Bang ,

Dong Ju Mun ,

Hoigi Seo ,

Abstract

Generative posterior sampling using diffusion models for inverse problems in imaging is enhanced through optimized scheduling of data consistency, classifier-free guidance, and stochasticity components via a time-varying control approach.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Generative posterior sampling using diffusion models has emerged as a dominant paradigm for solving inverse problems in imaging, which usually consists of three main components: data consistency (DC) guidance, classifier-free guidance (CFG) and stochasticity. While prior arts have focused on how to develop each or all components, less attention has given to how to schedule them, leading to heuristically fixed or partially adjusted suboptimal schedules. In this work, we argue that the interactions among all three components in terms of scheduling are crucial for significantly improved performance in solving inverse problems in imaging. Our analysis shows that aggressive CFG early in sampling conflict with DC guidance, while stochasticity brings the trajectory back to higher-probability regions. Based on these findings, we propose Triadic Dynamics Aware Posterior Sampling (TriPS), which reformulates posterior sampling as a time-varying control problem and optimizes schedules following a triadic trend of decreasing DC and stochasticity scales alongside increasing CFG scale. TriPS achieves this through two strategies: template-based search over functional priors for reliable baseline schedules, and Group Relative Policy Optimization (GRPO)-based reinforcement learning for more flexible temporal curves. Experiments demonstrate TriPS outperforms state-of-the-art baselines in data fidelity and perceptual realism.

View arXiv page View PDF Project page GitHub 18 Add to collection

Community

dongju1875

Paper author 1 day ago

•

edited 1 day ago

TriPS: Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems 🚀

Generative posterior sampling with diffusion and flow matching models has become a dominant paradigm for solving inverse problems in imaging, and it usually rests on three components: data consistency (DC) guidance, classifier-free guidance (CFG), and stochasticity. Prior arts focused on how to develop each component, but gave less attention to how to schedule them, leading to heuristically fixed or partially adjusted suboptimal schedules. We argue that the interactions among all three components, in terms of scheduling, are crucial for significantly improved performance.

💬 Core Idea:
Our analysis shows that aggressive CFG early in sampling conflicts with DC guidance, while stochasticity brings the trajectory back to higher-probability regions. Building on this, we reformulate posterior sampling as an optimization problem of time-varying schedules and jointly optimizes the three schedules along a triadic trend:

We identify the inherent conflict between DC guidance and CFG, formalized in Prop. 1: when CFG opposes the DC direction, a larger CFG scale hinders minimization of the measurement residual.
We reveal stochasticity as a regularizer that aligns sampling trajectories toward higher-probability regions, counteracting the off-manifold tendency of DC guidance and CFG.
This yields one clean rule: monotonic decrease in β(t) (DC), increase in λ(t) (CFG), and decrease in η(t) (stochasticity).

✨ TriPS realizes this trend through two complementary strategies. A template-based search over functional priors (linear, exponential, logarithmic) identifies reliable baseline schedules, and a GRPO-based reinforcement learning stage (Bernstein–Beta policy) captures more flexible temporal curves that transcend fixed functional forms. In effect, the sampler is told to enforce data consistency first and sharpen semantics later.

🎯 Why it matters:

It runs zero-shot on pretrained backbones (SD3.5-M, SD1.5), with no retraining.
Experiments demonstrate that TriPS outperforms state-of-the-art baselines in data fidelity and perceptual realism. On motion deblurring (FFHQ), TriPS reaches 31.20 dB PSNR against 28.80 for the next best, with LPIPS of 0.060 against 0.095, and stays best or second-best across SR ×8/×12, deblurring, and inpainting.
A hybrid reward that jointly optimizes distortion (PSNR) and perceptual metrics (LPIPS) lets GRPO navigate the perception-distortion trade-off on demand.

📄 ICML 2026 · arXiv:2605.26470
🔗 Project: https://mundongju.github.io/TriPS/ · Code: https://github.com/mundongju/TriPS