EquiReg: Equivariance Regularized Diffusion for Inverse Problems
Bahareh Tolooshams [email protected], Aditi Chandrashekar * [email protected], Rayhan Zirvi * [email protected], Abbas Mammadov [email protected], Jiachen Yao [email protected], Chuwei Wang [email protected], Anima Anandkumar [email protected] · 2025
Read EquiReg: Equivariance Regularized Diffusion for Inverse Problems by Bahareh Tolooshams [email protected], Aditi Chandrashekar * [email protected], Rayhan Zirvi * [email protected], Abbas Mammadov [email protected], Jiachen Yao [email protected], Chuwei Wang [email protected], Anima Anandkumar [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-22973.txt.
Introduction
Bahareh Tolooshams and co-authors keep diffusion-based image repair on the rails of plausibility, in a 2025 paper. Diffusion solvers restore damaged images by sampling what fits both the picture's prior and the measurements. The measurement step is usually fudged with a simple Gaussian guess, which can shove the sample off the manifold of real images. Their plug-in, EquiReg, fines paths that break the data's symmetries.
Symmetry error runs low for true images and high for fakes, a quirk of models trained with augmentation on symmetric data. Penalizing it steers sampling toward honest regions, speeding convergence. Gains peak when sampling steps are scarce, exactly where rivals degrade. Restoration tasks, linear and not, improved across the board.
Catalog record
| CREATOR | Bahareh Tolooshams [email protected], Aditi Chandrashekar * [email protected], Rayhan Zirvi * [email protected], Abbas Mammadov [email protected], Jiachen Yao [email protected], Chuwei Wang [email protected], Anima Anandkumar [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 102 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-22973
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
EquiReg: Equivariance Regularized Diffusion for Inverse Problems Bahareh Tolooshams [email protected], Aditi Chandrashekar * [email protected], Rayhan Zirvi * [email protected], Abbas Mammadov [email protected], Jiachen Yao [email protected], Chuwei Wang [email protected], Anima Anandkumar [email protected] Abstract Diffusion models represent the state-of-the-art for solving inverse problems such as image restoration tasks. Diffusion-based inverse solvers incorporate a likelihood term to guide prior sampling, generating data consistent with the posterior distribution. However, due to the intractability of the likelihood, most methods rely on isotropic Gaussian approximations, which can push estimates off the data manifold and produce inconsistent, poor reconstructions. We propose Equivariance Regularized (EquiReg) diffusion, a general plug-in framework that improves posterior sampling by penalizing trajectories that deviate from the data manifold. EquiReg formalizes manifold-preferential equivariant functions that exhibit low equivariance error for on-manifold samples and high error for off-manifold ones, thereby guiding sampling toward symmetry-preserving regions of the solution space. We highlight that such functions naturally emerge when training non-equivariant models with augmentation or on data with symmetries. EquiReg’s largest gains are under reduced sampling and measurement consistency steps, where many methods suffer severe quality degradation. By regularizing trajectories toward the manifold, EquiReg implicitly accelerates convergence and enables high-quality reconstructions. EquiReg consistently improves performance in linear and nonlinear image restoration tasks and solving partial differential equations. Our code is available at https://github.com/Anima-Lab/EquiReg . Reviewed on OpenReview: https://openreview.net/forum?id=3iaYKJLcyG 1 Introduction Inverse problems aim to recover an unknown signal {\bm{x}}^{*}\in\mathbb{R}^{d} from undersampled noisy measurements: {\bm{y}}={\mathcal{A}}({\bm{x}}^{*})+\bm{\nu}\in\mathbb{R}^{m}, (1) where {\mathcal{A}} is a known measurement operator, and \bm{\nu} is an unknown noise ( Groetsch, 1993 ) . Inverse problems are widely studied in science and engineering, with applications from medical imaging to astrophotography. Inverse problems are ill-posed, i.e., the inversion process can have many solutions; hence, they require prior information about the desired solution ( Kabanikhin, 2008 ) . In the Bayesian formulation, inference targets the posterior distribution p({\bm{x}}|{\bm{y}})\propto p({\bm{y}}|{\bm{x}})p({\bm{x}}) , where p({\bm{y}}|{\bm{x}}) is the likelihood of the measurements and p({\bm{x}}) is a prior describing the signal structure ( Stuart, 2010 ) . Examples of handcrafted priors include sparsity ( Donoho, 2006 ) and low-rankness ( Candès et al., 2011 ) . This paper focuses on methods that leverage unconditionally pre-trained score-based generative diffusion models as learned priors ( Ho et al., 2020 ; Song and Ermon, 2019 ) with applications in image restoration ( Chung et al., 2023 ) , medical imaging ( Chung et al., 2022a ) , and solving partial differential equations (PDEs) ( Huang et al., 2024 ; Yao et al., 2025 ) . These methods define a sequential noising process {\bm{x}}_{0}\sim p_{\mathrm{data}}\to{\bm{x}}_{t}\to{\bm{x}}_{T}\sim p_{T}({\bm{x}})\approx\mathcal{N}(\bm{0},{\bm{I}}) and a reverse denoising process parameterized by a neural network score \nabla_{{\bm{x}}_{t}}\log p_{t}({\bm{x}}_{t}) ( Vincent, 2011 ) . During sampling, these approaches incorporate gradient signals carrying likelihood information to solve inverse problems. Figure 1 : Equivariance Regularized (EquiReg) diffusion for inverse problems. (left) Manifold preferential equivariance (MPE) functions whose equivariance error is lower for on-manifold and higher for off-manifold data. (right) EquiReg regularizes the posterior sampling trajectory for improved performance. It penalizes off-manifold trajectories via MPE-based regularization. This is a schematic illustrating the motivation of our method; the depicted EquiLoss color gradient is qualitative rather than a direct measurement of manifold distance, with empirical equivariance error shown in Figure 4(a) . \mathcal{M}_{t} denotes a high-density region of p_{t} at diffusion time t , rather than its support, with \mathcal{M}_{0}=\mathcal{M} denoting the clean data manifold. Solving inverse problems with diffusion ( Zhang et al., 2025a ; Alkhouri et al., 2025 ) requires computing the conditional score \nabla_{{\bm{x}}_{t}}\log{p_{t}({\bm{x}}_{t}|{\bm{y}})} , decomposed into \nabla_{{\bm{x}}_{t}}\log{p_{t}({\bm{x}}_{t})}+\nabla_{{\bm{x}}_{t}}\log{p_{t}({\bm{y}}|{\bm{x}}_{t})} . This introduces challenges, as the likelihood score \nabla_{{\bm{x}}_{t}}\log{p_{t}({\bm{y}}|{\bm{x}}_{t})}=\nabla_{{\bm{x}}_{t}}\log{\int{p({\bm{y}}|{\bm{x}}_{0})p_{t}({\bm{x}}_{0}|{\bm{x}}_{t})\mathrm{d}{\bm{x}}_{0}}} is only computationally tractable when t=0 . To handle the likelihood for t>0 , many methods approximate the posterior p_{t}({\bm{x}}_{0}|{\bm{x}}_{t}) with the isotropic Gaussian distribution ( Zhang et al., 2025a ) , where the distribution expectation is computed using the optimal denoising score ( Robbins, 1956 ) . The Gaussian approximation can be inaccurate for complex distributions ( Figure 3 ), leading to errors in likelihood computation, especially with point estimations ( Chung et al., 2023 ) . Since the posterior expectation is a conditional expectation, a linear combination of all possible {\bm{x}}_{0} , it may lie off the data manifold even when individual samples remain on it. These issues are further amplified in latent diffusion models (LDMs), introducing artifacts ( Rout et al., 2023 ) . Prior work has attempted to address this challenge via projection-based ( He et al., 2024 ; Zirvi et al., 2025 ) or decoupled optimization strategies ( Zhang et al., 2025a ) , aimed at reducing the propagation of measurement errors during sampling. However, they rely on the isotropic Gaussian assumption, which can lead to failures on difficult tasks or at reduced sampling steps. While higher-order statistics reduce such errors ( Boys et al., 2024 ) , most approaches employ a Gaussian approximation for its efficiency, scalability, and simplicity ( Alkhouri et al., 2025 ) , particularly in the context of large-scale LDMs ( Peebles and Xie, 2023 ) . This raises a key question: how can we ensure that conditional diffusion samples remain on the data manifold under this approximation? Equivariance offers a natural mechanism to keep sampling trajectories close to the data manifold. We address this challenge with a regularization scheme that leverages equivariance to improve posterior sampling by guiding trajectories toward symmetry-preserving solution spaces. Prior work has enforced equivariance directly within generation or denoising processes ( Chen et al., 2023a ; Terris et al., 2024 ) , with extensions to probabilistic symmetries ( Bloem-Reddy et al., 2020 ) enabling sample efficiency ( Wang et al., 2024 ) . Figure 2 : EquiReg’s broad applicability. a-d) image restoration inverse problems and e) text-guided image generation, resulting in artifact reduction and more realistic generation. Here, EquiReg refers to our regularization being applied to the diffusion sampling method on the same row. Our approach differs as follows: rather than strictly enforcing equivariance within denoising architectures, which can hinder tasks requiring symmetry breaking ( Lawrence et al., 2025 ) , we employ equivariance as a plug-in regularizer to guide diffusion trajectories toward the data manifold. Our contributions. We propose Equivariance Regularized (EquiReg) diffusion, an equivariance-based regularization framework for solving inverse problems with diffusion models ( Figures 1 and 2 ). EquiReg leverages equivariance to regularize likelihood-induced errors during posterior sampling, guiding diffusion trajectories toward more consistent, on-manifold solutions. Crucially, it employs Manifold-Preferential Equivariant (MPE) functions, which discriminate on-manifold from off-manifold data by exhibiting low equivariance error in-distribution and higher error out-of-distribution. Figure 3 : Off-manifold posterior expectation. This impacts the likelihood score p_{t}({\bm{y}}|{\bm{x}}_{t})=\int{p({\bm{y}}|{\bm{x}}_{0})p_{t}({\bm{x}}_{0}|{\bm{x}}_{t})\mathrm{d}{\bm{x}}_{0}} computation achieved via isotropic Gaussian modelling of p_{t}({\bm{x}}_{0}|{\bm{x}}_{t}) . The right panel is a schematic 1D illustration of p_{t}({\bm{x}}_{0}|{\bm{x}}_{t}) for a fixed {\bm{x}}_{t} , with the curve being the (bimodal) posterior density. The Tweedie posterior mean {\bm{x}}_{0|t} is a value on the {\bm{x}}_{0} -axis, and the pink dotted segment connects the two modes to emphasize that {\bm{x}}_{0|t} falls between them in a low-density region. We formalize that a regularizer for diffusion posterior sampling should capture such a global property, and MPE functions provide a principled way to direct sampling toward plausible solutions. This design makes EquiReg architecture-agnostic: the regularizer operates independently of the diffusion model itself. With