Technical Summary: Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems

1. Problem Statement

Inverse problems in imaging aim to recover an unknown signal $x_0$ from noisy measurements $y = Ax_0 + n$, where $A$ is a known forward operator and $n$ is noise. While classical methods rely on hand-crafted priors, recent paradigms utilize generative priors from diffusion models and flow matching to solve these ill-posed problems.

Posterior sampling in this context typically relies on three core components to guide the reverse diffusion process:

1. Data Consistency (DC) Guidance: Enforces alignment between the estimated signal and the measurements ($Ax \approx y$).
2. Classifier-Free Guidance (CFG): Steers the sampling trajectory toward a conditioning text prompt to leverage semantic priors.
3. Stochasticity: Injects noise during the reverse process to maintain diversity and prevent mode collapse.

Existing works often treat these components with time-independent (constant) schedules or only partially adjust them (e.g., varying only DC). This leads to suboptimal performance because the interactions between these components are dynamic. Specifically, aggressive CFG early in the process can conflict with DC guidance, and excessive guidance without stochasticity can drive trajectories away from high-probability regions of the data manifold.

2. Methodology: TriPS

The authors propose TriPS (Triadic Dynamics Aware Posterior Sampling), a framework that reformulates posterior sampling as a time-varying control problem. The core insight is that the three components form a coupled "triadic system" requiring coordinated, time-dependent scheduling.

2.1 Analysis of Triadic Coupling Dynamics

The paper establishes three key theoretical and empirical findings:

• Early-Stage Guidance Conflict: There is an inherent conflict between DC guidance and CFG in the early, high-noise stages of sampling. High CFG scales ($\lambda(t)$) early on can counteract the direction of DC guidance, slowing the minimization of the measurement residual and causing semantic hallucinations.
• Stochasticity as a Regularizer: Stochasticity acts as a regularizer that counteracts the "off-manifold" phenomenon. While strong DC and CFG can push sampling trajectories toward lower-probability regions, appropriately scaled stochasticity ($\eta(t)$) pulls the trajectory back toward the high-probability data manifold.
• Triadic Scheduling Trend: Based on the above, the optimal strategy follows a specific trend:
  • DC Guidance ($\beta(t)$): Monotonically decreasing. High early to enforce structure, low late to avoid over-fitting noise.
  • CFG ($\lambda(t)$): Monotonically increasing. Low early to avoid conflict with DC, high late to refine semantics and details once the global structure is established.
  • Stochasticity ($\eta(t)$): Monotonically decreasing. High early to regularize the trajectory, low late to ensure precise convergence to the target distribution.

2.2 Optimization Framework

To realize these schedules, TriPS employs two complementary strategies:

1. Template-Based Schedule Search (TriPST):

   • Searches a discrete space of functional priors (Linear, Exponential, Logarithmic) that satisfy the monotonicity constraints.
   • Uses a grid search on a small calibration set to maximize a utility function balancing distortion (PSNR) and perceptual quality (LPIPS).
   • Provides a robust baseline and a warm-start for further optimization.

2. GRPO-Based Schedule Optimization (TriPSG):

   • Uses Group Relative Policy Optimization (GRPO), a reinforcement learning framework, to discover complex, non-functional temporal curves that transcend fixed templates.
   • Parameterization: Schedules are modeled as continuous curves using Bernstein polynomials, with coefficients sampled from learnable Beta distributions. This ensures the schedules remain within physically valid bounds.
   • Reward: A hybrid reward function combines distortion metrics (PSNR) and perceptual metrics (LPIPS, CLIP-IQA+, Q-Align) to navigate the perception-distortion trade-off.
   • Policy Update: The policy is updated using a clipped surrogate objective with group-wise reward standardization, avoiding the need for a value function.

3. Key Contributions

• Analysis of Triadic Coupling: The paper identifies and formalizes the conflict between early-stage CFG and DC guidance, and characterizes stochasticity as a necessary regularizer to maintain alignment with the data manifold.
• Triadic Scheduling Framework: Proposes a unified scheduling trend (decreasing DC/Stochasticity, increasing CFG) that aligns with the coarse-to-fine nature of diffusion sampling.
• Dual Optimization Paradigms: Introduces a two-stage optimization approach: a template-based search for robust baselines and a GRPO-based method for fine-grained, task-specific schedule adaptation.
• Generalizability: Demonstrates the framework's applicability across different generative backbones (Flow Matching and Diffusion Models) and various inverse problems (Super-Resolution, Deblurring, Inpainting).

4. Experimental Results

The authors evaluated TriPS on high-resolution benchmarks (FFHQ, DIV2K) across multiple linear inverse problems: Super-Resolution (8x, 12x), Motion Deblurring, Gaussian Deblurring, and Inpainting.

• Quantitative Performance:

  • TriPS (both TriPST and TriPSG variants) consistently outperformed state-of-the-art baselines (e.g., ReSample, FlowChef, FlowDPS, FLAIR, PSLD, DDPG) in both distortion metrics (PSNR, SSIM) and perceptual metrics (FID, LPIPS, KID, MUSIQ).
  • For example, in Super-Resolution 8x on FFHQ using Flow Matching, TriPSG achieved a PSNR of 28.55 and LPIPS of 0.107, surpassing the previous best (FLAIR) which had a PSNR of 28.88 but a significantly higher LPIPS of 0.123.
  • The method showed strong transferability; schedules optimized for SR 8x generalized effectively to SR 12x and Deblurring tasks.

• Qualitative Results:

  • Visual comparisons demonstrated that TriPS mitigates structural artifacts and semantic hallucinations common in baselines with fixed or poorly scheduled guidance.
  • The GRPO-optimized schedules (TriPSG) allowed for flexible control over the perception-distortion trade-off, enabling variants that prioritize sharp textures (perception-oriented) or structural fidelity (distortion-oriented).

• Ablation Studies:

  • Joint optimization of all three components (DC, CFG, Stochasticity) yielded superior results compared to optimizing only one or two components while fixing the others.
  • The GRPO-based approach successfully learned complex temporal curves that deviated from simple templates, proving the necessity of flexible schedule optimization.

5. Significance and Claims

The paper claims that the proposed TriPS framework provides a robust basis for adaptive control in generative modeling for inverse problems. By explicitly modeling the time-varying interactions between data consistency, semantic guidance, and stochasticity, the method achieves high-fidelity restoration with improved perceptual realism.

The authors emphasize that their work offers new theoretical insights into the dynamics of posterior sampling, specifically the role of stochasticity as a regularizer against guidance conflicts. They state that these findings have broad practical utility in fields such as medical imaging, scientific reconstruction, and remote sensing, while raising no ethical concerns regarding the technology itself. The work focuses on enhancing reconstruction accuracy without adverse societal consequences.