For decades, the standard way to recover a signal from incomplete or noisy measurements was to assume it was "sparse." If you knew an image could be represented by just a few non-zero coefficients in a wavelet transform, you could use the Lasso to reconstruct it.
However, a new paradigm has emerged. Instead of assuming a signal is sparse, we assume it lives on the manifold of a pretrained generative model. If a GAN or a Diffusion model has learned what "natural images" look like, we can use that knowledge as a powerful prior to solve inverse problems—from deblurring and inpainting to compressed sensing and phase retrieval.
The Manifold Constraint
The core intuition is a shift in search space. In traditional compressed sensing, we search the entire high-dimensional space $\mathbb{R}^n$ for a vector that matches our measurements. With a generative prior, we restrict our search to the range of a generator $G(z)$, where $z$ is a low-dimensional latent code.
As Paul Hand, Oscar Leong, and Vladislav Voroninski demonstrated in Phase Retrieval Under a Generative Prior, this approach can achieve near-optimal sample complexity. Because the latent space $k$ is much smaller than the signal dimension $n$, we need far fewer measurements to "pin down" the correct signal. Pei Peng and Shirin Jalali showed in Using auto-encoders for solving ill-posed linear inverse problems that for a model with latent dimension $k$, roughly $2k$ measurements are often sufficient for recovery, regardless of the model's internal complexity.
The Representation Error Bottleneck
While generative priors are powerful, they are not perfect. Every paper in this field highlights a recurring challenge: representation error.
A generative model is a closed system; it can only produce signals that fall within its learned distribution. If you try to reconstruct an image that contains a unique detail the generator never saw during training, the model will simply "hallucinate" the closest thing it knows. Manik Dhar, Aditya Grover, and Stefano Ermon describe this in Modeling Sparse Deviations for Compressed Sensing using Generative Models as a "performance ceiling." No matter how many measurements you take, you cannot exceed the fidelity of the generator itself.
To break this ceiling, researchers have developed hybrid methods:
- Sparse Deviations: Manik Dhar et al. proposed Sparse-Gen, which models a signal as $G(z) + s$, where $s$ is a sparse vector that captures the fine details the generator missed.
- Image Adaptation: In Image-Adaptive GAN Based Reconstruction, Shady Abu Hussein suggests fine-tuning the generator's weights at test time to better fit the specific measurements of the target image.
- Uncertainty Modeling: Jian Wang and colleagues introduced CS-BGM in Uncertainty Modeling in Generative Compressed Sensing, which adaptively expands the generator's range by treating network parameters as variables to be optimized.
Navigating the Non-Convex Landscape
Optimizing a latent code $z$ to match measurements $y$ is a non-convex problem. It is prone to getting stuck in local minima where the reconstructed image looks "natural" but doesn't actually match the data.
To solve this, Tianci Liu and Qi Lei developed Amortized Inverse Problem Optimization (AIPO). As detailed in Optimization for Amortized Inverse Problems, AIPO gradually ramps up the influence of the generative prior. By starting with a simple data-fitting objective and slowly introducing the complex prior, the optimizer is "guided" toward the global minimum, avoiding the spurious traps of the non-convex landscape.
From Point Estimates to Posterior Sampling
Most early work focused on finding a single "best" reconstruction (Maximum A Posteriori or MAP). However, inverse problems are often ill-posed, meaning many different signals could explain the same measurements.
Recent work has shifted toward posterior sampling. Instead of finding one image, we want to sample from the distribution of all likely images. Ajil Jalal and Alex Dimakis proved in Instance-Optimal Compressed Sensing via Posterior Sampling that posterior sampling is a near-optimal strategy for any signal distribution.
This is where Diffusion Models and Normalizing Flows shine. For instance, Gabriel Cardoso and Eric Moulines introduced MCGdiff in Monte Carlo guided Diffusion for Bayesian linear inverse problems, a method that uses Monte Carlo guidance to ensure diffusion-based reconstructions are mathematically consistent with the observed data. Similarly, Jianfeng Lu and Yao Xie leverage the speed of Consistency Models to perform Posterior sampling via Langevin dynamics based on generative priors, allowing for rapid exploration of diverse, high-quality solutions.
Theoretical Foundations
The field is also maturing theoretically. Zhaoqiang Liu and Jonathan Scarlett have provided rigorous guarantees for these methods. In The Generalized Lasso with Nonlinear Observations and Generative Priors, they showed that signals can be recovered even when the measurement process is non-linear or the data is corrupted by an adversary, provided the signal is "near" the range of the generative model. They further extended this logic to Generative Principal Component Analysis, showing that replacing the sparsity assumption with a generative prior allows for PCA with significantly fewer samples.