Skip to content

🖼️ Image Restoration

🧠 NeurIPS2026 · 29 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (78) · 📷 CVPR2026 (140) · 🔬 ICLR2026 (62) · 🧪 ICML2026 (21) · 🤖 AAAI2026 (12) · 🧠 NeurIPS2025 (29)

🔥 Top topics: Diffusion Models ×8 · Super-Resolution ×6 · Image Restoration ×5

A Principled Self-Referenced Early Stopping Approach for Deep Image Prior

To tackle Deep Image Prior's vulnerability to overfitting measurement noise due to over-parameterization, this paper establishes a self-referenced validation framework that derives three task-tailored early stopping algorithms—channel-similarity reference (CSR-ES), mask reference (MR-ES), and augmented-channel spectral bias reference (ACR-ES)—achieving near-oracle stopping accuracy without requiring ground-truth noise level estimates.

Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models

Under a shift-invariant Gaussian prior assumption, this paper derives closed-form spectral distributions for both zero-shot diffusion posterior samplers and the ideal posterior sampler, replacing ad-hoc hyperparameter heuristics with principled guidance weights optimized by minimizing the averaged Wasserstein-2 distance to the true posterior across image deblurring, inpainting, and accelerated MRI reconstruction.

Beyond Spatial-Domain Supervision: A Relation Constrained Space for Multi-Modal Image Fusion

Rather than treating source images as ideal fused images, RCS-Fusion learns a relational supervision space from frozen DINO/CLIP features and trains different fusion backbones with shared, complementary, and conflict-coordination constraints; its Transformer variant achieves 0.730 mAP in detection using fused M3FD images.

Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding

To tackle flicker-banding artifacts caused by temporal aliasing between rolling shutters and display brightness dimming, this paper presents Bricker—a dual synthetic-real bracketed RAW dataset constructed via physical ray-tracing simulation and automated multi-exposure capture—alongside BRACE, a multi-frame restoration network leveraging frequency-aware priors, multi-scale spatial cross-attention fusion, and a custom evaluation metric SFC.

Can We Model the Artifacts Explicitly? Disentangle Artifacts via Pairwise Edit Residuals

Reinterpreting image manipulation localization as a latent-variable inference problem, this paper introduces the EditGroup-45K dataset with 45K edit groups and proposes Pairwise Artifacts Learning (PAL) with marginal-matching negative sampling to explicitly disentangle manipulation artifacts from semantic content, providing an architecture-agnostic initialization that resists semantic shortcuts.

Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration

Addressing global prompt bottlenecks and discrete MoE routing instability in all-in-one restoration, Continuous Expert Assembly (CEA) synthesizes instance-conditioned low-rank routing bases and residual directions via a lightweight cross-attention hyper-adapter, enabling each spatial token to assemble continuous residual updates through dense signed dot products.

Control-Augmented Autoregressive Diffusion for Data Assimilation

Addressing the compounding error and prohibitive test-time optimization of autoregressive diffusion models in chaotic forecasting, this paper introduces CADA, an amortized stochastic optimal control framework that trains an offline controller to preview future observations and inject affine corrections into diffusion denoising sub-steps, enabling test-time-optimization-free, order-of-magnitude faster data assimilation with superior long-horizon stability.

Exact Posterior Score Estimation for Solving Linear Inverse Problems

This paper derives the closed-form exact posterior score for linear Gaussian inverse problems under general Gaussian interpolants, shows that posterior sampling reduces to anisotropic denoising at an operator-dependent shifted pivot, and proposes EPS, a denoising objective that preserves pretraining structure and achieves superior fidelity and perceptual quality without likelihood gradients or projections.

Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems

This paper demonstrates that flow map models implicitly parameterize a one-parameter family of two-time average denoisers via their lookahead parameter \(t\), which provably traces the optimal distortion-perception (DP) frontier for Gaussian targets and enables continuous, zero-shot DP traversal across general inverse problems when embedded into a Plug-and-Play framework.

GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation

By transposing 3D Gaussian Splatting into continuous RGB color space, GLUT models 3D lookup tables via a compact set of explicit learnable 3D Gaussian primitives and local affine mappings, achieving state-of-the-art color fidelity under 1 KB model size, multi-style blending, and training-free real-time local editing.

Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution

Addressing severe spatial activation heterogeneity induced by selective scan in Mamba super-resolution, this paper proposes Token-Group Quantization (TGQ) that statically predicts token clipping groups from pre-scan features without runtime statistics overhead, combined with a Frequency-Preserving Refinement (FPR) calibration objective, achieving state-of-the-art low-bit (W4A4/W2A2) restoration and a 2.23× real speedup on Jetson Orin Nano.

i-DEQ: A stable inertial deep equilibrium model for image restoration

i-DEQ integrates restarted inertial optimization into training and inference for an explicit-energy deep equilibrium model, achieving reconstruction quality close to ELDER with substantially shorter equilibrium-solving time in small image restoration experiments, but its theory establishes conditional stationarity complexity and practical training is not always stable.

Learning Where and What to Restore for Composite Image Restoration

CART allocates restoration compute through window-level spatial routing inside a U-Net, then controls image-level channel routing with task features shaped by multi-hot degradation supervision, achieving 29.95 dB average PSNR on CDD-11 while trading its speed advantage for more parameters and higher peak memory than MoCE-IR.

LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models

To tackle the unresolved trade-off between perceptual quality and fidelity as well as the failure of conventional IQA metrics for generative low-level vision, this paper introduces LL-Bench—comprising 16 degradation tasks, 28.9k restored images, and 152k expert preference pairs—and proposes LL-Score, a unified MLLM evaluator and reward model that jointly assesses restoration quality and hallucination probability.

LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution

LucidNFT addresses severe structural and semantic hallucinations in flow-based real-world super-resolution by introducing LucidConsistency (a degradation-invariant, inpainting-negated LR evaluator), decoupled group-wise reward normalization, and LucidLR (a 20k real-world degradation dataset).

MESS: Multi-Exposure Sequence Synthesis for Generalizable Image Enhancement

Addressing the severe cross-task generalization bottleneck and training data scarcity in image illumination enhancement, MESS uses camera-consistent exposure priors derived from RAW images to guide a diffusion-based generator, synthesizing realistic multi-exposure sequences from single in-the-wild 8-bit RGB images to build a 500K-pair dataset that dramatically elevates downstream generalization across low-light, exposure correction, and backlit benchmarks.

Multidimensional Observer Model and Perceptual Dimensions of Human Image Quality Assessment

The paper represents images as Gaussian distributions in a low-dimensional perceptual space and fits human choices through noisy candidate–reference distance comparisons; the CORnet-S Multi observer reaches 95% of its own above-chance performance at 3, 9, and 96 dimensions on BAPPS, PieAPP, and NIGHTS, respectively, suggesting task-dependent perceptual structures.

ODDR: One-Step Deshadow Diffusion via Reward Guidance

ODDR proposes an efficient one-step diffusion deshadowing framework free from real-world paired supervision, bootstrapping a single-step baseline (ODD) on physically synthesized shadows and refining it on real unpaired shadow images via ShadowReward—the first shadow-tailored reward model trained via ranked progressive synthetic degradations without human annotations.

PCDFusion: Proposal-Context-Detail Bayesian Rendering for Infrared-Visible Image Fusion

PCDFusion reformulates infrared-visible image fusion into a visible-anchored, three-stage Bayesian posterior rendering pipeline (Proposal-Context-Detail) that uses physical priors and learned log-evidence to adaptively resolve exposure failures and thermal contamination.

Principled Design of Diffusion-based Optimizers for Inverse Problems

This paper introduces OptDiff, a principled optimization framework for diffusion-based inverse problem solving that eliminates tedious hyperparameter tuning through tolerance-based noise schedule bounds and two-sided normalized gradients, enabling seamless integration of momentum and polynomial preconditioning for substantial inference speedups.

PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution

PRISM addresses the dual challenges of unreliable prior estimation and ambiguous local boundaries in severely degraded text super-resolution by combining Flow-Matching Prior Rectification (FMPR) with a Structure-guided Uncertainty-aware Residual Encoder (SURE) within a single-step diffusion backbone.

PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing

PRISM reformulates real-world dehazing into a physically grounded deep unfolding framework (PSAR) that jointly optimizes spatial atmospheric light, transmission, and scene radiance via closed-form proximal updates, complemented by an online non-uniform haze synthesis pipeline and selective self-distillation adaptation (SSDA) with physical residual-haze auditing.

Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing

Q-Probe is the first agentic IQA framework that scales image quality assessment to high-resolution scenarios by constructing the wavelet-injected Vista-Bench and establishing a progressive three-stage curriculum (Pairwise GRPO pre-alignment, context-aware SFT warm-up, and decoupled Post-RL) to eliminate the spurious "cropping implies degradation" bias.

RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement

To tackle flicker-banding degradations caused by temporal aliasing between rolling-shutter camera readout and display brightness modulation, RIFLE introduces the first physics-inspired banding simulation pipeline, a pixel-aligned real-world benchmark, and a one-step latent diffusion restoration framework guided by a flicker prior estimator and area-decoupled masked loss.

Single Shot HDR Recovery via a Video Diffusion Prior

By reformulating single-shot high dynamic range (HDR) image reconstruction as conditional video generation followed by lightweight fusion, this paper leverages a video diffusion prior to generate a motion-free, monotonically increasing exposure bracket and fuses it with a compact UNet into high-fidelity HDR radiance maps.

SP\(^3\): Spherical Priors for Plug-and-Play Restoration

SP\(^3\) introduces a gradient-free Plug-and-Play (PnP) restoration algorithm leveraging Spherical Encoders (SE) as manifold priors, alternating noisy spherical manifold projections with closed-form data-fidelity steps under Half-Quadratic Splitting to match diffusion-level perceptual quality while running 33–630\(\times\) faster with anytime restoration capability.

SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation

This paper introduces artifact prominence as a human-centered target for super-resolution defect evaluation, builds the 3,935-mask SR-Prominence crowdsourced dataset suite across four domains, discovers that 48.2% of in-lab binary artifacts in DeSRA go unnoticed by most viewers, and establishes a threshold-free correlation benchmark alongside a pseudo-GT evaluation protocol requiring no additional crowdsourcing.

When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution

Addressing the fidelity loss and structural hallucinations caused by the VAE compression bottleneck in latent diffusion transformers, PGSR preserves pre-VAE pixel evidence from the low-resolution observation and deploys a dual-stage guidance framework that steers latent flow trajectories and grounds frozen VAE decoder rendering.

WorldSR: Harnessing World Knowledge Search for Grounded Image Super-Resolution

WorldSR introduces an entity-grounded image super-resolution framework that decomposes low-resolution inputs into semantic entities, retrieves and filters aligned real-world reference images via web and local search with memory caching, and fuses them through a Dual-LoRA one-step diffusion model to eliminate generative hallucinations.