Robust Ensemble Guidance for Scientific Inverse Problems

Zixiang Li; Wei Wang; Yunchao Wei; Yao Zhao; Yue Song

cs.LGOct 4, 2026

Robust Ensemble Guidance for Scientific Inverse Problems

Authors: Zixiang Li, Wei Wang, Yunchao Wei, Yao Zhao, Yue Song

Organizations: Beijing Jiaotong University · Tsinghua University

Abstract

Ensemble guidance combines pretrained diffusion priors with black-box forward models to solve inverse problems without differentiating through the physical simulator. However, observation coordinates with large predictive spread or extreme residuals can dominate the ensemble correction, degrading reconstruction accuracy. We show that two simple modifications, weighting and clipping, substantially improve this correction. Our method, Robust Ensemble Guidance (REG), uses ensemble predictive spread to balance observation scales and adaptively clips standardized residuals to limit the influence of extreme discrepancies. Both operations reuse existing particles and forward predictions, requiring no additional denoiser or forward-model evaluations. Under a local linear Gaussian model, we derive conditions for reduced one-step estimation risk, bound the influence of individual observation coordinates, and characterize when these benefits persist with finite ensembles. Experiments on Navier-Stokes inversion, black-hole imaging, and acoustic full-waveform inversion demonstrate improved reconstruction over the underlying ensemble solver. In particular, REG increases black-hole reconstruction PSNR by 6.2-8.2 dB across three observation regimes and reduces Navier-Stokes reconstruction error by 26.4% in a matched-budget comparison. These findings highlight the importance of observation heterogeneity and residual influence in designing reliable generative solvers for scientific inverse problems.

Figures & tables

Figure 1: Overview of Robust Ensemble Guidance. Forward predictions from a denoised particle ensemble determine coordinate weights and a shared clipping threshold. The resulting bounded residuals are transported back to the particle space to form the next correction.

Figure 2: Controlled tests of weighting, clipping, and the conditional one-step risk criterion.

Subsampling	$\times 2$			$\times 4$			$\times 8$
Method / $\sigma_{y}$	$0$	$1$	$2$	$0$	$1$	$2$	$0$	$1$	$2$
EKI	0.577 -0.6pt	0.609 -0.6pt	0.673 -0.6pt	0.579 -0.6pt	0.669 -0.6pt	0.805 -0.6pt	0.852 -0.6pt	0.940 -0.6pt	1.116 -0.6pt
DPS-fGSG	1.687 -0.6pt	1.612 -0.6pt	1.454 -0.6pt	1.203 -0.6pt	1.209 -0.6pt	1.200 -0.6pt	1.246 -0.6pt	1.221 -0.6pt	1.260 -0.6pt
DPS-cGSG	2.203 -0.6pt	2.117 -0.6pt	1.746 -0.6pt	1.175 -0.6pt	1.133 -0.6pt	1.114 -0.6pt	1.186 -0.6pt	1.204 -0.6pt	1.218 -0.6pt
DPG	0.325 -0.6pt	0.408 -0.6pt	0.466 -0.6pt	0.322 -0.6pt	0.361 -0.6pt	0.454 -0.6pt	0.596 -0.6pt	0.591 -0.6pt	0.846 -0.6pt
SCG	0.908 -0.6pt	0.928 -0.6pt	0.966 -0.6pt	0.869 -0.6pt	0.926 -0.6pt	0.929 -0.6pt	1.260 -0.6pt	1.284 -0.6pt	1.347 -0.6pt

Table 1: Navier–Stokes relative

\ell_{2}

error (

\downarrow

). Cells show mean above (s.d.); bold indicates the lowest mean in each column.

$J$	EnKG	REG	Change
512	$0.2752\pm 0.1088$	$\mathbf{0.2690}\pm 0.0890$	$-2.3\%$
1024	$0.2029\pm 0.1250$	$\mathbf{0.1493}\pm 0.0463$	$-26.4\%$
2048	$0.1156\pm 0.0385$	$\mathbf{0.1000}\pm 0.0314$	$-13.4\%$

Table 2: Matched ablation (

\times 2

\sigma_{y}=0

). Relative

\ell_{2}

error: mean

\pm

s.d.; bold: lowest mean per comparison. Changes are relative to EnKG at the same

J

Figure 3: Navier–Stokes reconstructions (top) and absolute errors (bottom) for the same test case, with shared color scales across methods.

Obs. time	$\mathbf{3\%}$				$\mathbf{10\%}$				$\mathbf{100\%}$
Method	PSNR $\uparrow$	Blur $\uparrow$	$\widetilde{\chi}^{2}_{cp}$	$\widetilde{\chi}^{2}_{\log ca}$	PSNR $\uparrow$	Blur $\uparrow$	$\widetilde{\chi}^{2}_{cp}$	$\widetilde{\chi}^{2}_{\log ca}$	PSNR $\uparrow$	Blur $\uparrow$	$\widetilde{\chi}^{2}_{cp}$	$\widetilde{\chi}^{2}_{\log ca}$
SMILI	18.51 -0.6pt	23.08 -0.6pt	1.478 -0.6pt	4.348 -0.6pt	20.85 -0.6pt	25.24 -0.6pt	1.209 -0.6pt	21.788 -0.6pt	22.67 -0.6pt	27.79 -0.6pt	1.878 -0.6pt	17.612 -0.6pt
EHT-Imaging	21.72 -0.6pt	25.66 -0.6pt	1.507 -0.6pt	1.695 -0.6pt	22.67 -0.6pt	26.66 -0.6pt	1.166 -0.6pt	1.240 -0.6pt	24.28 -0.6pt	28.57 -0.6pt	1.251 -0.6pt	1.259 -0.6pt
DPS	24.20 -0.6pt	30.83 -0.6pt	8.024 -0.6pt	5.007 -0.6pt	24.36 -0.6pt	30.79 -0.6pt	13.052 -0.6pt	6.614 -0.6pt	25.86 -0.6pt	32.94 -0.6pt	8.759 -0.6pt	5.456 -0.6pt
LGD	22.51 -0.6pt	28.50 -0.6pt	15.825 -0.6pt	12.862 -0.6pt	22.08 -0.6pt	27.48 -0.6pt	10.775 -0.6pt	13.375 -0.6pt	21.22 -0.6pt	26.06 -0.6pt	13.239 -0.6pt	13.233 -0.6pt
RED-diff	20.74 -0.6pt	26.10 -0.6pt	6.713 -0.6pt	9.128 -0.6pt	22.53 -0.6pt	27.67 -0.6pt	2.488 -0.6pt	4.916 -0.6pt	23.77 -0.6pt	29.13 -0.6pt	1.853 -0.6pt	2.050 -0.6pt

Table 3: Black-hole imaging over 100 images. Cells show mean above (s.d.); bold and underline mark the best and second-best PSNR and blurred PSNR (dB).

Figure 4: Selected black-hole reconstructions of one case at $3\%$ , $10\%$ , and $100\%$ observation time, with PSNR (dB). EnKG and REG share observations; other baselines retain native realizations.

Method	Relative $\ell_{2}\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$	Data misfit $\downarrow$
Adam	0.333 (0.086)	9.968 (2.083)	0.305 (0.120)	115.14 (52.10)
DPS	0.250 (0.154)	14.111 (6.820)	0.491 (0.161)	155.08 (92.17)
LGD	0.244 (0.024)	12.288 (0.889)	0.341 (0.047)	258.47 (26.40)
DiffPIR	0.204 (0.129)	16.113 (6.962)	0.554 (0.191)	88.53 (56.91)
DAPS	0.201 (0.103)	14.914 (4.184)	0.321 (0.067)	111.13 (71.33)
PnP-DM	0.259 (0.075)	11.983 (2.269)	0.431 (0.073)	308.84 (26.34)

Table 4: Acoustic FWI results. Mean (s.d.). Bold marks the best mean in each column.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: NS reconstructions and absolute errors at $\times 2$ , $\sigma_{y}=0$ . EnKG and REG use $J=2048$ ; numbers are relative $\ell_{2}$ errors of the displayed reconstructions.

Figure 6: REG reconstructions of NS id0 across downsampling factors $\times 2/4/8$ and noise levels $\sigma_{y}=0/1/2$ , with $J=2048$ . All panels share a vorticity scale.

Figure 7: Selected black-hole reconstructions: id0 at $100\%$ , id3 at $3\%$ , and id7 at $3\%$ observation time. Numbers are aligned PSNR (dB); evaluation follows Appendix B.3 .

Figure 8: Additional black-hole targets and REG reconstructions at $3\%$ and $100\%$ observation time, with $J=256$ . Numbers are aligned PSNR (dB).

Figure 9: FWI reconstructions and absolute errors for id2, id5, and id7, with $J=64$ . Velocity and error panels use shared color scales; numbers are relative $\ell_{2}$ errors.

Explore similar work

May 29, 2025cs.LG

EquiReg: Equivariance Regularized Diffusion for Inverse Problems

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

Bahareh Tolooshams, Aditi Chandrashekar, Rayhan Zirvi +4

University of Alberta · Alberta Machine, Intelligence Institute (Amii) · Neuroscience and Mental Health Institute, University of Alberta +3

Bayesian Inverse Problems Inverse Problem Image Restoration Regularization Manifolds

Ask PDF

Jan 29, 2026cs.LG

Inverse Problems Conditioned on Observation Ensembles: Applications and Methods

We introduce a new multivariate statistical problem that we refer to as the Ensemble-conditioned Inverse Problem (EIP). The aim of EIP is to invert for an ensemble that is distributed according to the pushforward of a prior under a forward process. In high energy physics (HEP), this is related to a widely known problem called unfolding, which aims to reconstruct the true physics distribution from observations that are distorted by detector effects. The EIP also arises in full waveform inversion (FWI) and inverse imaging with unknown priors. We propose non-iterative inference-time methods that construct posterior samplers based on a new class of conditional generative models, which we call ensemble inverse generative models. For the posterior modeling, these models additionally use the ensemble information contained in the observation set on top of single observations. Unlike existing methods, our proposed methods avoid explicit and iterative use of the forward model at inference time via training across several sets of truth-observation pairs that are consistent with the same forward model, but originate from a wide range of priors. We empirically demonstrate that this training procedure can implicitly encode the likelihood model, enabling direct posterior inference for unseen priors to some degree. We benchmark the proposed method on several synthetic and real datasets in inverse imaging, HEP, and FWI. Our code is available at https://github.com/ZhengyanHuan/EIP.

Zhengyan Huan, Camila Pazos, Martin Klassen +3

Department of Electrical and Computer Engineering, Tufts University, Medford, Massachusetts · The NSF AI Institute for Artificial Intelligence and Fundamental Interactions · Department of Physics and Astronomy, Tufts University, Medford, Massachusetts +1

Inverse Problem Posterior Sampling Full Waveform Inversion Generative Models

Ask PDF

May 15, 2026cs.LG

Learning Normalized Energy Models for Linear Inverse Problems

Generative diffusion models can provide powerful prior probability models for inverse problems in imaging, but existing implementations suffer from two key limitations:

(i)

the prior density is represented implicitly, and

(ii)

they rely on likelihood approximations that introduce sampling biases. We address these challenges by introducing a new energy-based model trained for denoising with a covariance-based regularization term that enforces consistency across different measurement conditions. The trained model can compute normalized posterior densities for diverse linear inverse problems, without additional retraining or fine tuning. In addition to preserving the sampling capabilities of diffusion models, this enables previously unavailable capabilities: energy-guided adaptive sampling that adjusts schedules on-the-fly, unbiased Metropolis-Hastings correction steps, and blind estimation of the degradation operator via Bayes rule. We validate the method on multiple datasets (ImageNet, CelebA, AFHQ) and tasks (inpainting, deblurring), demonstrating competitive or superior performance to established baselines.

Nicolas Zilberstein, Santiago Segarra, Eero Simoncelli +1

Rice University, Houston, TX, USA · Flatiron Institute, New York, NY, USA · New York University, New York, NY, USA

Bayesian Inverse Problems Denoising Diffusion Probabilistic Model Inverse Problem Diffusion Models Regularization Posterior

Ask PDF