What Makes the Posterior Score So Hard to Compute?

According to the paper, "Existing methods either steer a fixed pretrained denoiser with approximate measurement-matching corrections, or train a conditional restoration model that abandons the denoising structure of the prior." The core problem is that diffusion models provide the score of the unconditional data distribution, but solving an inverse problem requires the score of the posterior distribution p(x|y). The unconditional score knows what natural images look like, but it doesn't know which image is consistent with the measurement y. Prior work approximated this correction using gradients of the likelihood, which are only accurate under restrictive assumptions about the noise level and measurement operator. The authors reported that their key insight is that the posterior score can be expressed exactly as the unconditional score plus a correction term that is itself a denoising problem. This correction term can be computed using the same pretrained denoiser, without any retraining or approximation. The result is a sampling procedure that provably converges to the true posterior.

How Does This Compare to the Leading Approximate Methods?

To understand the practical impact, consider the three most widely used steering approaches: Diffusion Posterior Sampling (DPS), Denoising Diffusion Null-space Model (DDNM), and Manifold Constrained Gradient (MCG). Each makes a different approximation to avoid computing the exact posterior score. The table below compares them head-to-head with the exact method.

Method	Approximation Error	Requires Retraining	Works for Any Linear A?	Sampling Speed
DPS (Chung et al., 2023)	High – uses heuristic gradient scaling	No	Yes	Slow – requires backprop through denoiser
DDNM (Wang et al., 2023)	Medium – assumes null-space consistency	No	Only for orthogonal A	Fast
MCG (Chung et al., 2022)	High – manifold assumption often violated	No	Yes	Slow
Exact (This Paper)	Zero – mathematically exact	No	Yes – any linear A	Comparable to DPS
Verdict: Exact method dominates all approximate methods on accuracy, generality, and retains the plug-and-play advantage.

Which Applications Will Benefit First?

The paper explicitly lists linear inverse problems including inpainting, super-resolution, deblurring, computed tomography (CT), and magnetic resonance imaging (MRI). These are domains where measurement operators are linear and well-characterized. According to the authors, the method "works for any linear measurement operator" without modification. This is a direct threat to companies like Subtle Medical (AI-enhanced MRI) and HeartVista (automated cardiac MRI), which currently use task-specific conditional models. Those models are expensive to train and maintain for each new scanner or protocol. An exact plug-and-play method could render them obsolete. However, the method does not yet handle nonlinear inverse problems (e.g., phase retrieval in coherent diffraction imaging). The authors note this as a limitation. For nonlinear problems, approximate steering methods will remain relevant for the foreseeable future.

What Are the Computational Costs of Exactness?

The exact posterior score correction requires solving an additional denoising problem at each sampling step. The authors reported that this increases per-step computation by approximately 2x compared to DPS. However, because the correction is exact, the total number of sampling steps can be reduced while maintaining fidelity. The paper reports that with 50 steps, the exact method achieves better reconstruction quality than DPS with 200 steps, making the wall-clock time competitive. This is a critical detail for deployment. In medical imaging, where a single MRI scan can generate gigabytes of data, reconstruction time matters. The paper's claim that exactness can be achieved without a net slowdown, if the step count is optimized, will be tested by replication studies. According to the paper, the method "enables high-quality reconstruction with as few as 50 sampling steps."

1. By Q1 2027, the Hugging Face Diffusers library will adopt the exact posterior score method as the default inverse problem solver, deprecating DPS.
2. By Q2 2027, at least two FDA-cleared AI medical imaging companies (Subtle Medical, HeartVista) will announce partnerships to integrate exact posterior score reconstruction into their workflows.
3. By Q3 2027, the number of arXiv papers citing this method will exceed 100, making it the dominant approach for linear inverse problems with diffusion models.

• Insight 1: The exact posterior score formula is a mathematical identity, not an approximation. This means it cannot be improved upon for linear inverse problems—it is the theoretical ceiling.
• Insight 2: The computational cost argument (2x per step but fewer steps) is the paper's most leveraged claim. If it holds, deployment barriers vanish.
• Insight 3: The paper does not address nonlinear inverse problems, leaving a clear research gap for the next generation of methods.
• Insight 4: The timing is perfect: diffusion models are now standard, but the lack of exact inference has limited their practical use in high-stakes applications like medical imaging.
• Insight 5: The paper's authors are anonymous on arXiv, but the mathematics is clean enough that the result is likely correct. The community will quickly verify and adopt.