In traditional signal processing, restoring a degraded signal—whether it is a blurry image, a reverberant voice recording, or a noisy seismic scan—usually requires a hand-crafted "prior." These priors, such as sparsity or smoothness, act as mathematical guesses about what a clean signal should look like. However, hand-crafted priors are often too simple to capture the rich complexity of real-world data.
A new research direction replaces these rigid constraints with generative diffusion priors. Instead of telling the computer that a signal should be "smooth," researchers train a diffusion model to learn the entire statistical distribution of clean data. Once trained, this model acts as a high-dimensional map, providing a "score function"—a vector field that points from noisy, unrealistic states toward the manifold of realistic, clean signals.
The Manifold as a Guide
The core intuition behind this approach is that all "valid" signals (like clean speech or clear medical images) live on a low-dimensional manifold within a much larger space of possible noise. Saito et al. argue that models trained exclusively on clean data can outperform supervised models because they focus on the target distribution rather than the specific transformation of the degradation.
When we encounter a degraded signal, we are essentially looking at a point that has been knocked off this manifold. By using a pre-trained diffusion model, we can iteratively "steer" the degraded signal back toward the manifold. As Wang et al. and colleagues demonstrate in the context of seismic imaging, the learned score function guides the reconstruction process, providing a more expressive and "smarter" regularizer than traditional methods.
Diffusion Posterior Sampling
To solve an inverse problem, we don't just want a realistic signal; we want a realistic signal that is consistent with our observations. This is typically achieved through Diffusion Posterior Sampling (DPS).
In this framework, the reconstruction process is a tug-of-war between two forces:
- The Prior: The diffusion model pushes the signal to look like the training data (e.g., clean speech).
- The Likelihood: A physics-based observation model ensures the reconstruction matches the recorded data.
Welker et al. show how this works for audio dereverberation: by leveraging a clean-speech prior and a known room impulse response, the system "steers" the generation toward an anechoic version of the signal. For image reconstruction, Zhang et al. utilize Tweedie’s formula—a mathematical tool that allows a denoiser to estimate the clean signal at any stage of the process—to move toward high-probability regions of the image manifold while maintaining consistency with linear measurements.
Solving the "Blind" Problem
The most challenging inverse problems are "blind," meaning the degradation process itself (like the specific acoustics of a room) is unknown. Recent work has shown that diffusion priors are powerful enough to solve for both the signal and the degradation simultaneously.
In the BUDDy framework, Moliner et al. treat blind dereverberation as a joint optimization. The system uses the diffusion prior to define what speech should sound like, while simultaneously optimizing a physics-informed model of the room's acoustics. Because the diffusion model has such a strong internal representation of speech, it can effectively distinguish between the target voice and the acoustic artifacts, as noted by Moliner et al. in their work on joint acoustics estimation.
This principle extends beyond audio. Xu et al. apply similar logic to separate overlapping voices in unknown environments, and Wu et al. demonstrate that these unsupervised methods can actually exceed the perceptual quality of supervised models, which often struggle to generalize to new, unseen environments.
Cross-Domain Versatility
Because the diffusion prior is "unconditional"—meaning it only cares about the final output distribution—it can be plugged into diverse domains without redesigning the core architecture:
- Medical Imaging: Liu et al. use dual-domain diffusion priors to remove metal artifacts from CT scans without needing paired "clean" and "artifact" training data.
- Biological Science: Shabeeb et al. reconstruct 3D protein structures from undersampled Cryo-EM images, using the prior to fill in missing biological signals.
- Robotics and Sensing: Yang et al. enhance sparse radar point clouds by using a diffusion prior trained on high-resolution LiDAR data, guiding the radar toward physically plausible structures.
By shifting the focus from "learning the mapping" to "learning the manifold," this research direction provides a robust, unsupervised path for high-fidelity signal recovery across the physical sciences.
Key Foundational Work
Unsupervised vocal dereverberation with diffusion-based generative models (2022) — contributed 2 concept elements to this direction.
Diffusion Posterior Sampling for Informed Single-Channel Dereverberation (2023) — contributed 1 concept elements to this direction.
A prior regularized full waveform inversion using generative diffusion models (2023) — contributed 1 concept elements to this direction.