⚡ electrical engineering

Dual Ascent Diffusion for Inverse Problems

This paper introduces Dual Ascent Diffusion, a novel dual ascent optimization framework that leverages diffusion model priors to solve ill-posed inverse problems with superior image quality, noise robustness, speed, and observation fidelity compared to existing state-of-the-art methods.

Original authors: Minseo Kim, Axel Levy, Gordon Wetzstein

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Minseo Kim, Axel Levy, Gordon Wetzstein

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a puzzle, but someone has taken a photo of the puzzle pieces, smudged the photo with grease, and then cut out a huge chunk of it. Your goal is to guess what the original, perfect puzzle looked like. In the world of science and engineering, this is called an inverse problem. It's like trying to figure out what a person looked like before they walked through a foggy, distorted mirror.

For a long time, computers have been good at guessing the missing pieces, but they often "hallucinate" (make up details that aren't there) or produce blurry, inaccurate results.

This paper introduces a new method called DDiff (Dual Ascent Diffusion) that acts like a smarter, more disciplined puzzle solver. Here is how it works, using simple analogies:

1. The Problem: The "Guessing Game"

Current methods use powerful AI models (called diffusion models) that have learned what "normal" images look like. Think of these models as an artist who has seen millions of photos and knows exactly how a face, a building, or a tree should look.

When solving a puzzle, these methods try to combine two things:

The Clues: What the blurry, damaged photo actually shows.
The Artist's Intuition: What the image should look like based on the AI's training.

The problem is that existing methods often struggle to balance these two. They might listen too much to the clues and produce a noisy mess, or listen too much to the artist and invent details that weren't in the original photo (hallucinations).

2. The Solution: A "Three-Person Team"

The authors created a new framework called DDiff that treats the problem like a negotiation between three specific roles, constantly checking each other's work. They use a mathematical strategy called Dual Ascent (inspired by a method called ADMM) to keep everyone in sync.

Think of the process as a team of three people trying to restore a damaged painting:

Person A (The Data Keeper): Their only job is to make sure the final image matches the blurry clues (the measurements). They constantly say, "Hey, this part doesn't look like the photo we were given!"
Person B (The Art Expert): Their job is to make the image look realistic and sharp, using the AI's knowledge of how images should look. They say, "That edge looks too jagged; let's smooth it out to look like a real photo."
Person C (The Referee): This is the new, crucial addition. Person C holds a "dual variable" (a scorecard). They watch Person A and Person B. If Person A and Person B disagree, Person C adjusts the tension between them. They ensure that the Art Expert doesn't invent fake details and that the Data Keeper doesn't get stuck in the noise.

3. How They Work Together (The Dance)

Instead of just taking a guess and hoping for the best, the team takes turns making small adjustments in a loop:

The Art Expert cleans up the image to make it look realistic.
The Data Keeper tweaks the image to make sure it fits the blurry clues.
The Referee checks the "gap" between the two. If the image is drifting too far from the clues, the Referee pulls it back. If it's too noisy, the Referee lets the Art Expert smooth it out.

The paper proves mathematically that if you keep doing this dance, the team will eventually stop arguing and agree on a single, perfect solution. This is called convergence to a fixed point.

4. Why It's Better

The paper claims DDiff is superior to previous methods for three main reasons:

It's More Honest: It creates images that match the original blurry clues much more closely. In the paper's tests, the "residual error" (the difference between the guess and the actual clue) was closer to zero. This means it invents fewer fake details.
It Handles Noise Better: If the original photo is very dirty or noisy (like a photo taken in a heavy storm), DDiff doesn't panic. It stays robust and doesn't get confused by the static, whereas other methods might produce a distorted mess.
It's Faster: Because the team works efficiently together, DDiff reaches a high-quality result in fewer steps than other methods. It's like solving the puzzle in half the time.

5. Real-World Tests

The authors tested this "three-person team" on various difficult tasks, such as:

Super-resolution: Turning a tiny, blurry image into a large, sharp one.
Inpainting: Filling in huge missing chunks of an image (like a 128x128 square missing from a face).
Deblurring: Fixing images that are smeared because the camera was moving.
Phase Retrieval: A complex physics problem where you have to reconstruct an image from wave patterns (often used in microscopy).

In all these tests, DDiff produced sharper, cleaner images with fewer artifacts than the current state-of-the-art methods.

Summary

In short, DDiff is a new way to use AI to fix broken images. Instead of letting the AI guess freely, it puts the AI in a structured "team" with a strict referee. This ensures the final result is both realistic (it looks like a real photo) and accurate (it actually matches the blurry clues we started with), even when the clues are very noisy or incomplete.

Technical Summary: Dual Ascent Diffusion for Inverse Problems

Problem Statement

Inverse problems involve estimating an unknown signal or image $x$ from noisy and corrupted observations $y$ , governed by a forward operator $A(\cdot)$ . These problems are ubiquitous in fields ranging from astrophysics and medical imaging to protein structure determination. They are typically ill-posed, necessitating the use of priors to constrain the solution space.

While modern approaches utilize pretrained diffusion models as powerful priors, existing methods for solving Maximum-a-Posteriori (MAP) problems or sampling from the posterior face significant limitations. Many rely on computational approximations to combine the likelihood of the observation model with the diffusion prior. These approximations often lead to inaccurate samples, suboptimal reconstruction quality, or a failure to faithfully represent the observations (i.e., high residuals), which can manifest as "hallucinations" in the generated solution. Furthermore, many state-of-the-art methods struggle with high levels of measurement noise and suffer from computational inefficiency due to the accumulation of errors or the need for supplementary Markov Chain Monte Carlo (MCMC) steps.

Methodology: Dual Ascent Diffusion (DDiff)

The authors propose DDiff, a new framework for solving MAP problems using pretrained diffusion models. Unlike methods that attempt to provably sample from the posterior, DDiff focuses on deriving an optimization strategy that accurately and efficiently solves the MAP problem. The method is inspired by the Alternating Direction Method of Multipliers (ADMM) but introduces a novel dual ascent structure specifically adapted for diffusion models.

Core Algorithmic Components

The DDiff framework iterates through three primary updates within each step, alternating between the primal variables ( $x, z$ ) and the dual variable ( $u$ ):

x-update (Data Fidelity): Instead of solving a constrained least-squares problem exactly, DDiff employs a linearized ADMM approach. It performs a single gradient descent step on the data fidelity term $-\log p(y|x)$ , allowing the method to be directly applicable to any differentiable forward model $A(\cdot)$ , whether linear or nonlinear.
$x \leftarrow v - \gamma \nabla_v \|y - A(v)\|_2^2, \quad \text{where } v = z - u$
z-update (Denoising): This step replaces the naive application of a diffusion model as a denoiser. The authors observe that standard score models $s_\theta(x, t)$ are only accurate on the diffusion manifold at time $t$ . To address the fact that the variable $x+u$ generally does not lie on this manifold, DDiff constructs a recursive variable $x_t$ that anchors the denoising process. The update involves a DDIM-style reverse diffusion step where the "predicted $x_0$ " is the current primal estimate $x$ , and the dual variable $u$ is re-scaled and added to maintain consistency with the signal level.
$z \leftarrow \frac{1}{\sqrt{\bar{\alpha}_t}} (x_t + (1-\bar{\alpha}_t)s_\theta(x_t, t))$
where $x_t$ is updated via a DDIM step incorporating the dual variable $u$ .
Dual Update: The dual variable $u$ is updated to enforce the constraint $x=z$ by accumulating the residual error between the primal variables:
$u \leftarrow u + x - z$

Latent Extension

The framework is extended to LatentDDiff for use with Latent Diffusion Models (LDMs). This variant operates on compact latent representations to reduce memory and compute demands. It supports two modes for enforcing data consistency:

Latent-DC: Enforces consistency directly in the latent space, requiring back-propagation through the decoder.
Pixel-DC: Decodes the latent variable to pixel space, applies the data consistency correction, and re-encodes. The authors note that Pixel-DC yields consistently higher-quality reconstructions.

Key Contributions

Dual Ascent Framework for Diffusion Priors: The paper introduces a novel optimization strategy that integrates dual ascent (inspired by ADMM) with diffusion models. This approach avoids the approximations common in existing posterior sampling methods.
Fixed-Point Convergence Proof: The authors establish the first fixed-point convergence proof for a diffusion-based posterior optimization method. They demonstrate that under bounded denoiser and gradient assumptions, the sequences of primal and dual variables converge to a fixed point $(x^\star, z^\star, u^\star)$ , even without assuming a convex prior.
Improved Robustness and Fidelity: The method is designed to produce solutions that more faithfully model the observations, evidenced by residuals closer to zero compared to state-of-the-art methods. This reduces the likelihood of hallucinations introduced by the generative prior.
Efficiency: DDiff achieves faster inference times by eliminating the need for supplementary MCMC equilibration steps (required by methods like DAPS) and reducing the number of neural function evaluations (NFEs) required for high-quality results.

Experimental Results

The authors evaluated DDiff on a wide range of linear and nonlinear inverse problems, including super-resolution, inpainting, Gaussian and motion deblurring, phase retrieval, nonlinear deblurring, and high-dynamic-range (HDR) imaging. Experiments were conducted on FFHQ and ImageNet datasets.

Quantitative Performance: DDiff outperformed state-of-the-art baselines (including DAPS, DMPlug, DCDP, DPS, DiffPIR, and DDRM) on the majority of tasks. It achieved superior scores in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Notably, it achieved the lowest residual errors, indicating better data consistency.
Robustness to Noise: The method demonstrated significantly greater robustness to high levels of measurement noise. While some baselines showed rapid degradation in PSNR as noise increased, DDiff maintained stable performance. In high-noise scenarios (e.g., phase retrieval with $\sigma=0.3$ ), DDiff recovered coherent global structures and semantics where other methods produced corrupted artifacts.
Efficiency: In terms of time efficiency, DDiff achieved better perceptual quality with fewer NFEs compared to DAPS. It required roughly half the sampling time for linear tasks and one-sixth the time for nonlinear tasks to achieve comparable or superior results.
Ablation Studies: Experiments confirmed that both the dual variable update and the specific noising step (Eq. 12 in the paper) are critical. Removing the dual variable (emulating DiffPIR) or the noising step resulted in significant performance drops, particularly in terms of high-frequency artifacts and data consistency.

Significance and Claims

The paper claims that DDiff represents a significant advancement in solving inverse problems with diffusion priors by shifting from posterior sampling approximations to a robust dual-ascent optimization framework. The authors emphasize that their method:

Faithfully models observations: By minimizing residuals, it reduces the "hallucination" effect often seen in generative inverse problem solvers.
Is computationally efficient: It removes the bottleneck of MCMC steps and reduces the computational burden of gradient calculations.
Is theoretically grounded: The provided convergence proof offers a theoretical guarantee for the stability of the method, a feature lacking in many prior diffusion-based optimization approaches.

The authors conclude that DDiff is a practical solution for large-scale, complex inverse problems, particularly in scenarios where measurement noise is prominent, such as low-dose CT reconstruction. Future work is suggested to extend the framework to higher-dimensional data like 3D volumes and video.