🖼️ Image Restoration

🎞️ ECCV2026 · 73 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (1) · 📷 CVPR2026 (135) · 🔬 ICLR2026 (61) · 🧪 ICML2026 (21) · 🤖 AAAI2026 (10) · 🧠 NeurIPS2025 (26)

🔥 Top topics: Image Restoration ×20 · Super-Resolution ×17 · Diffusion Models ×10 · Adversarial Robustness ×5 · Self-Supervised Learning ×4

Lumina-OmniLV: A Unified Multimodal Framework for General Low-Level Vision

Built on Lumina-Next generative priors, OmniLV combines separate text and image encoding, projection-addition of visual exemplars, and a co-trained shallow condition adapter to cover four low-level vision categories, with strong task alignment and several perceptual results but no across-the-board advantage over specialists.

Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging

Rather than redesigning the denoising network, this work estimates CCD-specific spatial response and readout noise from calibration frames and synthesizes training pairs from high-SNR stacked proxies, enabling a standard U-Net to reach average injected-source FLUX-SNR values of 39.91 and 39.87 dB on real MuSCAT-3 and cross-instrument MuSCAT-4 observations, respectively.

QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration

QualiTeacher turns teacher pseudo-label quality into a control condition for the student, then constrains quality ordering and cropped consistency so that imperfect supervision supports controllable real-world restoration, reducing RTTS dehazing FADE from 0.824 for CORUN with Colabator to 0.590.

Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion

Stream-DiffVSR combines four-step diffusion denoising, flow-aligned historical output guidance, and temporal decoding to restore video frame by frame without future frames, achieving REDS4 LPIPS of 0.099 and 0.328-second time-to-first-frame at 720p on an RTX 4090, although it does not yet provide conventional real-time HD video processing.

SparkVSR: Interactive Video Super-Resolution via Sparse Keyframe Propagation

SparkVSR lets users restore a few keyframes and propagate their visual priors under the motion constraints of the low-resolution video, reaching MUSIQ 68.88 and DOVER 0.6212 on MovieLQ with PiSA-SR references while exposing a guidance control for choosing between greater fidelity and richer generated detail.

Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution

StrSR proposes a joint framework of asymmetric adversarial distillation and frequency-domain distribution matching. By replacing the DiT discriminator with a CLIP-ConvNeXt discriminator, it avoids model collapse, and uses a Frequency Distribution Loss (FDL) to suppress grid-like periodic artifacts inherent in DiT. It achieves high-quality single-step real-world image super-resolution on DiT architectures for the first time.

λSplit: Self-Supervised Content-Aware Spectral Unmixing for Fluorescence Microscopy

λSplit learns cellular structural priors with a hierarchical VAE and remixes predicted fluorophore concentration maps through known spectra for training without unmixed ground truth; on simulated CellAtlas data with 5 ms exposure and 32 bands, it achieves 34.03 dB PSNR versus 28.91 dB for RLU, the strongest comparator by PSNR in that setting.

Calibrated Harmonic Overlaid Implicit Neural Representations for Multi-Dimensional Data

CHOIR replaces the layer-wise function composition of deep periodic INRs with β-gated additive harmonic superposition, then uses the 1/f power-law prior of natural images to calibrate both the frequency distribution and the output amplitude of every neuron, enabling stable periodic INR training at 12–15 layers and surpassing SIREN, FINER, FreSh, LRTFR, and CRNL on 2D image fitting, NeRF novel view synthesis, and hyperspectral / multispectral / video missing-data completion and mixed-degradation restoration.

RL-AWB: Deep Reinforcement Learning for Auto White Balance Correction in Low-Light Night-time Scenes

RL-AWB delegates illuminant estimation to the nighttime gray-pixel estimator SGP-LRD and uses a SAC policy trained on few images to adjust the gray-pixel percentage and aggregation exponent per image, achieving a median angular error of 1.99° when trained on LEVI and tested on NCC.

Experts-Guided Unbalanced Optimal Transport for ISP Learning from Unpaired and/or Paired Data

EGUOT leaves the deployed ISP network unchanged and combines a content anchor, unbalanced optimal transport, and color/structure/frequency experts during training, reaching 21.69 dB PSNR with a fixed Restormer in unpaired ZRR experiments while also supporting paired supervision.

From Local Windows to Adaptive Candidates via Individualized Exploratory: Rethinking Attention for Image Super-Resolution

IET lets each token independently filter and explore attention candidates from a local starting point, then reuses similar neighbors for channel fusion, improving Urban100 PSNR at four-times upscaling from PFT's 28.20 to 28.43 dB at the same 1.26T FLOPs.

OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution

OARS recasts the post-training of real-world image super-resolution (Real-ISR) as a process alignment: a multimodal reward model, COMPASS, is first trained to score the LR→SR transition itself — fidelity preservation plus perceptual gain, with the fidelity constraint tightened or relaxed according to input quality — after which the diffusion backbone Qwen-Image-Edit is aligned by three progressive online RL stages (cold-start flow matching, full-reference RL, non-reference RL), raising MUSIQ from 68.57 to 71.41 and LIQE from 3.81 to 4.30 on RealSR / DIV2K / RealSet80 while PSNR / SSIM stay essentially unchanged.

Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera Color

CCM-MLP preserves two-dimensional white-point chromaticity and uses a lightweight MLP to predict an image-wide color correction matrix, reducing mean angular color reproduction error from 4.56° with 2-CCM to 3.60° on the Pixel lightbox test set, primarily addressing modern illuminants away from the Planckian locus.

PASDiff: Physics-Aware Semantic Guidance for Joint Real-world Low-Light Face Enhancement and Restoration

PASDiff uses a training-free multi-objective energy guidance to solve low-light enhancement and blind face restoration jointly inside a single unconditional diffusion trajectory: on the physical side, an inverse-intensity-weighted spatially varying exposure constraint and a Retinex reflectance anchor stabilize illumination and chromaticity, while on the structural side an AdaIN statistic alignment distills only high-frequency structure from an off-the-shelf face restoration prior and explicitly strips away its intrinsic lighting and color bias, surpassing cascaded and joint baselines on both synthetic FFHQ and the self-built WildDark-Face benchmark (PSNR 22.59, face recognition accuracy 71.43%).

RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution

Treating the low-resolution image as a semantic anchor, the paper fine-tunes Qwen3-VL 8B with GRPO into an LR-conditioned, interpretable reward model (RefReward-SR) backed by the first large-scale LR-conditioned preference dataset (RefSR-18K), and then uses that reward as the core signal in a GRPO fine-tuning of the diffusion SR model C-FLUX — reaching 85.0% in-domain agreement with human annotators, on par with the 84.7% agreement among the annotators themselves.

Self-Evolving Agentic Image Restoration via Deliberate Planning and Intuitive Execution

SEAR discovers restoration tool sequences through long-horizon search and stores successful trajectories as verified memory shortcuts, achieving 22.1332 dB PSNR on MiO100 Group B while reducing average tool calls from 16.75 without memory to 8.15.

SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration

SLER-IR maps degradation features onto a unit hypersphere, uses cosine similarity to select independent experts at successive encoder-decoder layers, and guides restoration with global content and local degradation priors, achieving an average PSNR/SSIM of 31.73/0.928 across five tasks, exceeding MoCEIR by 1.15 dB/0.009.

Stokes-Informed Diffusion for Robust Linear Polarization Estimation

GenPolar reformulates the ill-posed RGB-to-polarization inverse problem as channel-wise Stokes prediction, supervises polarization angles in reliable regions, and uses one-step distillation to support VAE encoder adaptation, improving both polarization strength and angle estimation on three test groups.

TaskTok: Delving into Task Tokens for Task-driven Image Restoration

TaskTok proposes a task-driven image restoration (TDIR) framework. It leverages the property of 1D tokenizers where tokens are specialized by index position to encode different visual attributes. Through a learnable token switch and a lightweight token refiner, TaskTok selectively restores only the token subset most relevant to downstream tasks (classification, segmentation, or detection), substantially improving downstream task performance while drastically reducing computational overhead.

TEASR: Training-Efficient Any-Step Diffusion Transformer for Real-World Image Super-Resolution

TEASR performs self-adversarial distillation between two competing trajectories inside a single diffusion transformer, removing the need for a teacher model or a discriminator, and pairs it with decoupled timestep conditioning and time-aware rectification so that one ×4 super-resolution model supports any number of sampling steps from 1 to 30 without retraining.

Tuning Real-World Image Restoration at Inference: A Test-Time Scaling Paradigm for Flow Matching Models

ResFlow-Tuner turns a flow matching backbone such as FLUX.1-dev into a real-world image restorer (a SwinIR reference image plus a VLM caption are concatenated with the noisy latent into one sequence that conditions the model through self-attention), and adds a training-free test-time scaling stage at inference: the deterministic ODE trajectory is perturbed inside an equivalent SDE family to spawn N candidates, a multi-step partial denoising estimator cheaply previews each candidate's final image, and a rank-based ensemble of Aesthetic / CLIPScore / ImageReward keeps the top-M branches, converting controllable extra compute into a large lead on no-reference perceptual metrics.

Unified Removal of Raindrops and Reflections: A New Benchmark and A Novel Pipeline

The paper introduces unified removal of raindrops and reflections (UR3) and the real-captured RDRF dataset, combining initial restoration, dual-condition diffusion, and fidelity-assisted decoding in DiffUR3 to achieve 29.41 dB PSNR, 0.9372 SSIM, and 0.0813 LPIPS on RDRF-testing.

UniStitch: Unifying Semantic and Geometric Features for Image Stitching

UniStitch encodes sparse keypoints into grids aligned with image features and uses adaptive expert fusion to drive global-to-local stitching warps, reaching 25.07 dB / 0.813 on UDIS-D while reducing high-resolution warping memory through low-resolution TPS and B-spline reconstruction.

CogSENet: Blind Image Deblurring with Blur-Conditioned Semantic Routing and Explicit Frequency Fusion

CogSENet reformulates blind image deblurring from a passive pixel regression into an active, semantically aligned reconstruction process. By mimicking the active visual scanning, retinal functional differentiation, and focal adaptation of the eagle visual system, it proposes three core designs: a Semantic-Driven State Space Module (SDSSM), a Bi-Frequency Fusion Block (BFFB), and joint modulation based on a Continuous Blur Field (CBF) and CLIP semantics. With only 8.9M parameters, it outperforms state-of-the-art (SOTA) methods such as EVSSM and FFTformer on the GoPro, HIDE, and RealBlur datasets.

Fabric Image Demoiréing Benchmark from Synthesis to Restoration

This paper presents the first systematic study on the fabric image demoiréing problem. It introduces PRISM, a physical imaging chain simulation-based residual injection synthesis framework (including the first fabric moiré benchmark with 16,050 image pairs), and FaDeNet, a conservative restoration network specifically designed for the spectrally entangled characteristics of fabrics, significantly outperforming existing screen demoiréing methods in terms of PSNR, SSIM, and LPIPS.

Fidelity- and Perception-Aware Local Implicit Attention for Arbitrary-Scale Image Super-Resolution

FPLIA proposes a dual-stream framework that fuses the fidelity features of a regression backbone and the perceptual features of a diffusion model via asymmetric bidirectional cross-attention (FPAM) and pixel-wise adaptive selection (FPSM), simultaneously achieving high fidelity and high perceptual quality in ASISR.

FlowDec: Temporal Conditional Flow Decorruptor for Robust Continuous Vision-Language Navigation

FlowDec proposes an image decorruption framework based on conditional flow matching. By utilizing a hybrid temporal conditioning strategy and action centroid-guided filtering, it enhances robustness to multiple visual corruptions without updating the underlying VLN navigation backbone. It achieves relative navigation success rate gains of 25.33% and 9.38% on R2R-CE and RxR-CE, respectively, while running 3-8x faster during inference compared to diffusion-based TTA methods.

FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring

FMA-Net++ proposes a non-recurrent sequence-level framework based on Hierarchical Bidirectional Aggregation (HRBA) blocks. By introducing an Exposure Time-aware Modulation (ETM) module to inject frame-level exposure information into features, and combining it with an exposure-aware Flow-Guided Dynamic Filtering (FGDF) to jointly model motion- and exposure-varying degradations, the method achieves SOTA accuracy and inference speed on the VSRDB task.

Frames2Residual: Spatiotemporal Decoupling for Self-Supervised Video Denoising

To resolve the fundamental conflict between noise independence constraints and spatial texture recovery in self-supervised video denoising, F2R decouples learning into blind temporal consistency estimation and non-blind spatial texture refinement via a recorruption strategy, achieving state-of-the-art restoration on sRGB and raw videos.

Freqformer: Image-Demoiréing Transformer via Effective Frequency Decomposition

Freqformer introduces a smoothness-preserving, anti-aliasing frequency decomposition that splits moiré patterns into local high-frequency textures and scale-robust low-frequency color shifts, processed via asymmetric crop/resize branches and unified by a learnable Frequency Composition Transform (FCT), setting a new state-of-the-art across multiple high-definition demoiréing benchmarks with only 6.065M parameters.

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

FreqOrtho-SR proposes two core modules: Frequency-Guided LoRA Mixture-of-Experts (FreqMoE) and Orthogonal Gradient Projection (OGP). Driven by adaptive expert routing based on FFT degradation features and orthogonal constraints on pixel-semantic subspaces, it achieves degradation-adaptive real-world image super-resolution within a single-step diffusion model framework, obtaining optimal or sub-optimal trade-offs between fidelity and perceptual quality across multiple benchmarks.

Frozen CLIP Priors for Robust Self-Supervised Poisson Inverse Problems

For photon-limited Poisson inverse problems, this work presents an ADMM-inspired unrolled plug-and-play solver that decouples closed-form data-consistency updates from a parameter-efficient frozen CLIP multi-scale prior, achieving robust self-supervised reconstruction via measurement-domain GR2R recorruption and equivariant virtual acquisitions without clean ground truth.

HNDiff: Haze-Noise Diffusion for Image Dehazing

Addressing detail loss in regression-based dehazing under dense haze and fidelity degradation in standard Gaussian diffusion, HNDiff embeds the atmospheric scattering model into the diffusion process via haze-aware noise scheduling and dual-estimator reverse sampling, serving as an effective plug-and-play latent prior generator.

HOIMask: Towards Generative Masked Modeling for Human Object Interaction Generation

Addressing error accumulation and physical penetration artifacts in diffusion-based methods, HOIMask presents the first generative masked modeling framework in discrete space for full-body HOI motion synthesis, leveraging a 2D HOI VQ-VAE alongside inference-time contact-aware reconstruction guidance to produce highly coherent and physically plausible interactions.

Learning to Balance: Decoupled Siamese Diffusion Transformer for Reference-Based Remote Sensing Image Super-Resolution

To resolve the fundamental trade-off between reference texture over-reliance and under-utilization in remote sensing RefSR, DS-DiT decouples low-resolution structural priors from reference texture interactions within the attention mechanism, integrates patch-level adaptive weighting, and introduces training-free inference autoguidance to achieve high fidelity and realistic detail recovery under large scaling factors.

Leveraging Phase Information to Boost Unrolled Network Learning for Image Deblurring

Addressing the vulnerability of conventional unrolled deblurring networks to noise and their neglect of Fourier phase structure, this paper derives closed-form LMMSE estimators for degraded phase and amplitude and unrolls them into UPADNet, achieving state-of-the-art restoration with remarkable noise and low-data robustness.

Linear Fusion MultiDiffusion for Fast Training-Free Spherical Panorama Generation

This paper introduces LF-MultiDiffusion, which extends MultiDiffusion to arbitrary linear projection operators by reformulating latent fusion as a regularized least-squares problem solved via a matrix-free Krylov solver, synthesizing seamless high-quality 360° panoramas with only 14 perspective views while achieving a 15.36× speedup over the best training-free baseline.

Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

Live Avatar presents an algorithm-system co-designed framework featuring noisy KV cache (History Corrupt) and Timestep-forcing Pipeline Parallelism (TPP), enabling a 14B-parameter diffusion model to perform stable, real-time streaming generation at 45.2 FPS beyond 10,000 seconds on 5 H100 GPUs.

LogicIR: Logic Gate Networks for Image Restoration

LogicIR is the first logic gate network (LGN) specifically designed for image restoration. Through a fully logic-gate UNet architecture, a differentiable bit decoding layer, and an Index Shuffling cross-group communication mechanism, it achieves competitive restoration quality on denoising, deblocking, and deraining tasks with operations (BOPs) far lower than BNN and LUT methods, demonstrating the feasibility of pure logic gate operations in image restoration.

Perceiving Better Moments: Cover Frame Reselection and Enhancement for Live Photos with the Live2K Dataset

Addressing the severe quality gap in resolution, color fidelity, and dynamic range when reselecting cover frames in Live Photos, this paper constructs the first dual-device real-world benchmark Live2K and develops LPENet, a unified baseline integrating temporal-spatial multi-frame fusion, cross-attention color modulation, and patch-level texture transfer.

Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising

Spire models page-level presentation slide personalization (PSP) as an inverse planning problem. By applying discrete structural perturbations to golden slides, it constructs self-supervised "denoising" signals to train Critic and Planner agents (7B-level) to cooperatively and iteratively refine design specifications under RL. Without relying on specific executors, it infers the user's implicit design intent and significantly outperforms strong GPT-based baselines in both visual similarity and VLM-as-Judge scores.

Pixel Ignores, Superpixel Sees: Adverse Weather Image Restoration via Semantic-Center SSM

Addressing the issue where conventional rigid-grid SSM scans cross semantic boundaries and introduce degradation noise, this paper proposes the Semantic-center guided State space model for Restoration (SSR), constraining SSM recursive state evolution within perceptually coherent superpixels coupled with region-level gating, achieving unified all-weather restoration SOTA with only 7.05M parameters and 54.27G FLOPs.

ProSR: Semantic-Prototype-Guided Discrete Modeling for Physically Consistent SAR Super-Resolution

Addressing structural smearing and mode interpolation induced by smooth-approximation diffusion models in SAR super-resolution, ProSR decouples low-frequency semantics from high-frequency residuals via self-supervised discrete quantization, and constrains cross-attention using target, shadow, and clutter scattering prototype maps to achieve physically consistent reconstruction.

Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling

Addressing the acute shortage of historical high-resolution global reanalyses for 0.1° weather forecasting, BaguanHR shifts the paradigm from coarse-to-fine model transfer to variable-wise super-resolution data synthesis, establishing power-law scaling effects and surpassing IFS-HRES across 85% of lead times within 72 hours.

Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation

WINRO is a training-free and model-agnostic framework that discovers "winning noise tickets" from a precomputed null-prompt motion dictionary and refines them via KL-regularized optimization, effectively overcoming semantic drift in text-to-motion diffusion models.

Robust Self-Supervised Cross-Modal Super-Resolution against Real-World Misaligned Observations

RobSelf tackles the dual challenges of spatial misalignment and ground-truth absence in real-world multi-modal imaging through an online self-supervised model that jointly optimizes a weakly-supervised feature translator and a content-aware reference filter, delivering high-resolution and high-fidelity predictions without requiring training data, ground-truth supervision, or pre-alignment.

Spectral Gradient Orthogonalization Improves Differentially Private Training at Scale

Addressing the spectral mismatch where isotropic noise in DP-SGD overwhelms low-rank gradient energy in vision models, this paper introduces polar decomposition via Newton-Schulz iteration as a zero-cost post-processing step, uncovering an SNR phase transition where orthogonalization delivers massive accuracy gains under large batches and high model capacity.

Stabilizing Deep Reconstruction Operators with Contractive Anchoring

Addressing the pervasive "peak-and-collapse" instability where deep denoisers in iterative Plug-and-Play (PnP) and RED reconstruction degrade abruptly after reaching peak performance, this paper proposes an adaptive contractive anchoring framework that monitors local operator expansiveness and adaptively blends a lightweight contractive anchor in closed form, eliminating divergence without modifying or retraining black-box denoisers.

Stochastic Optimal Control Sampling for Diffusion Inverse Problems

By formulating the reverse denoising process as a controlled dynamical system, this paper derives a closed-form single-step stochastic optimal control update with bounded terminal penalties, bypassing full-trajectory second-order Hessian computation while adaptively balancing data fidelity and generative prior preservation.

The Devil Is in the Dark Pixels: Toward Brightness Bias-Robust Denoising

Reveals that MSE-trained image denoisers suffer from severe brightness bias due to signal-dependent residual inflation compounded by monotonically increasing Jacobian norms, and presents Brightness Bias-Robust Denoising (BBRD)—a plug-and-play loss with zero additional parameters or inference cost that simultaneously boosts performance across all brightness bands.

UHD-MFF: Shattering Barriers in Multi-Focus Ultra-High-Definition Image Fusion via Learnable Lookup Tables

This paper proposes the first 4K ultra-high-definition multi-focus image fusion dataset, UHD-MFF (1,950 pairs of 3840×2160 images), and designs a scale-decoupled learnable lookup table framework, UMF-LUT. It uses a low-resolution C-LUT for coarse region-level decision-making and a high-resolution D-LUT for edge-level refinement. With only 0.008M parameters, it achieves state-of-the-art fusion quality while enabling real-time 4K inference at 90fps, allowing deployment on mobile devices.

YeTI: You Only Need Two Noisy Images for Real-World sRGB Noise Generation

YeTI trains a structure-noise disentangling autoencoder plus a one-step conditional diffusion model from only two burst noisy images of the same scene; at generation time a single noisy image suffices to synthesize fresh real-world sRGB noise without clean references or camera metadata, reaching 0.0256 KLD / 0.1108 AKLD on the SIDD validation subset and providing effective training data for self-supervised denoisers.

340 FPS Reflection-free Video from Spikes Modulated by a Rapidly Rotating Polarizer

A rapidly rotating polarizer modulates spike-camera measurements, and reliability-weighted physical separation followed by residual refinement reconstructs reflection-free video at a nominal 340 FPS output cadence, achieving 27.0598 dB PSNR on the authors' synthetic test set.

BWAFDA: Block-wise Weighted Attention Fusion with Detail-aware for No-Reference Image Quality Assessment

Instead of directly averaging features over an entire image, BWAFDA preserves localized distortion evidence through dual-stream representations, mask-guided multi-scale aggregation, and class-token gating, achieving 0.777 PLCC and 0.723 SRCC on LIVEFB, absolute gains of 0.042 and 0.084 over the reported SHDIQA baseline.

CLDefocus: Physically Grounded Compound-Lens Defocus Blur Synthesis

Addressing the lack of optical diversity and physical realism in existing defocus deblurring datasets, this paper introduces a scalable synthesis pipeline combining Debye CZT wave-optics propagation, depth-aware occlusion handling, and linear-space ISP simulation to produce CLDefocus, a benchmark of 700 compound lenses that substantially improves cross-device deblurring generalization.

Degradation-Agnostic Clarity Learning for Unpaired Image Dehazing

Addressing severe semantic hallucinations caused by the cross-domain semantic gap in unpaired image dehazing, this paper proposes Degradation-Agnostic Clarity Learning (DCL), which formulates clarity extraction as multi-task Pareto optimization over adversarially mined degradations and integrates an Enhance-to-Resist self-supervised objective for detail recovery.

Difficulty-Conditioned Attribute-Specific Restoration for Low-Light Image Enhancement

Addressing spatial degradation heterogeneity in low-light imagery, DCASR decouples restoration in HVI space using a Global Intensity Modulator for coherent brightness adjustment and a DCT-driven Structural Refinement Modulator for localized detail recovery, complemented by a two-stage privileged teacher-student diffusion strategy for residual-free inference.

Diffusion-based dual-view reflection removal

Addressing the severe ill-posedness of single-image reflection removal and the vulnerability of multi-frame sequence methods to reflection-corrupted flow estimation, this paper presents a dual-view diffusion framework featuring directed frame-aware spatial attention to prevent cross-view noise leakage and soft smartphone IMU pose embeddings to disentangle reflection layers via physical parallax.

D²R²OSR: Degradation-Disentangled Representation for Real-World Omnidirectional Image Super-Resolution

Addressing the compounded challenges of fisheye capture degradations and Equirectangular Projection (ERP) geometric distortions in omnidirectional imaging, D²R²OSR introduces a dual-branch network coupled with a Perspective Projection Representation (PPR) that losslessly decouples viewport representations, achieving SOTA real-world ODI-SR performance with superior computational efficiency.

Event-driven Motion Deblurring via Trajectory-based Kernel Reconstruction

To overcome the severe ill-posedness of non-uniform motion deblurring, this paper estimates dense pixel-wise motion trajectories directly from event streams via a differentiable alignment objective, constructs physically grounded spatially varying point spread function (PSF) operators, and solves the joint deblurring problem using an ADMM-inspired deep unfolding network that achieves state-of-the-art restoration quality with low latency.

Extreme Face Super-Resolution through Identity Fitting and Decoupling

Addressing severe identity loss and pervasive facial hallucination in extreme degradation regimes (8×/16× scaling factors), IDFSR masks corrupted low-resolution facial regions to eliminate misleading cues, injects landmark-warped reference features via style-adaptive normalization alongside ArcFace identity embeddings into diffusion attention layers, and achieves unprecedented identity consistency and perceptual fidelity via lightweight test-time identity embedding optimization.

Fast and Accurate Image Restoration with Rank Enhanced Linear Attention

Addressing the severe performance collapse of linear attention in image restoration caused by its low-rank attention map, this paper proposes Rank Enhanced Linear Attention (RELA) to restore full rank via lightweight depthwise convolution, and develops LAformer—a hardware-efficient Vision Transformer devoid of softmax and window shifting—achieving SOTA restoration accuracy and impressive generative scalability with linear complexity.

FreeMEF: A Flexible Multi-Exposure Fusion Transformer for Arbitrary Number of Frames

FreeMEF proposes a "there and back again" two-stage paradigm: first, a recurrent state space module is utilized to progressively aggregate multi-exposure features of arbitrary frame counts into a global representation, which then guides reference frame restoration via extremist-aware hybrid attention. This significantly suppresses ghosting artifacts and enhances dynamic range while supporting flexible inference for 2/3/5 frames.

Hybrid-LUT: Channel-Aware Hybrid Lookup Table and Filtering for Efficient Image Restoration

To eliminate the heavy on-chip SRAM consumption caused by triplicating lookup tables (LUTs) across RGB channels, Hybrid-LUT proposes an asymmetric YUV framework that applies a multi-band LUT branch with pixel-wise adaptive fusion to the luminance (Y) channel while using lightweight mean filtering for the chrominance (UV) channels, reducing LUT storage by two-thirds and surpassing existing SOTA by 0.63 dB CPSNR on real-world denoising benchmarks with only 421 KB of storage.

Integrated Forward–Inverse Network for Lensless Image Reconstruction

To tackle severe multiplexing and calibration mismatch caused by large-footprint PSFs in lensless cameras, IFIN interleaves differentiable forward projections with learnable inverse updates across an encoder–decoder hierarchy, while optimizing a shift-variant PSF field end-to-end to set new state-of-the-art reconstruction benchmarks.

Learn to See the Unseen in Low-light Spike Streams

To overcome the severe ill-posedness of reconstructing images from sparse and noisy spike streams under extreme low-light and high-speed conditions, this paper proposes Diff-SPK, a diffusion-based framework that couples generative priors with spike-derived structural constraints using an Enhanced Texture from Inter-spike Interval (ETFI) representation, tailored condition encoding, and full-trajectory feature fusion.

LUCE: Constrained Curve-Domain Guidance for Training-Free Low-Light Enhancement with Hue-Preserving Decoupling

Addressing the instability and chromatic distortions caused by competing pixel-space attribute gradients in training-free diffusion-based enhancement, LUCE reformulates sampler-side guidance as a constrained curve-domain luminance control problem with hue-preserving decoupling, driving a frozen DDPM sampler with a single scalar luminance energy.

SLAIR: Structured Latent Flow Matching for All-in-One Image Restoration

SLAIR reformulates all-in-one image restoration as deterministic latent flow matching, utilizing a classification-guided separation to enlarge the manifold margin between degraded and clean latents, bridge-noise regularized ODE transport, and wavelet high-frequency decoding to achieve state-of-the-art fidelity and efficiency across 8 degradation types without explicit degradation labels.

Spatiotemporal Flux Probing for Single-Photon Videography

This paper introduces a spatiotemporal flux probing theory that discards the conventional reliance on local space-time windows or isolated pixels, estimating the continuous intensity spectrum directly from extremely sparse asynchronous photon streams via global 3D Fourier probing functions and constant false alarm rate (CFAR) detection, enabling joint recovery of ultrafast scene motion and high-frequency illumination flicker at 10× to 30× lower photon budgets along with velocity-selective video refocusing.

TextFace: Compositional Text-Guided Identity Preserving Face Synthesis for Face Recognition

TextFace introduces a dual-conditioned latent diffusion framework integrating compositional textual attribute control with score-space identity blending, decoupling extrinsic style variations from intrinsic identity traits to overcome intra-identity collapse in synthetic face recognition datasets.

TPCNet: A Low-Light Image Enhancement Network Inspired by Triple Physical Constraints

Addressing the limitations of conventional Retinex algorithms that assume ideal Lambertian reflection while neglecting specular components and impose physical constraints in fragile image space, TPCNet formulates a Triple Physical Constraints (TPC) theory derived from Kubelka-Munk imaging principles and implements it in implicit feature space with an efficient Cross-Guided Transformer, achieving state-of-the-art enhancement across 10 benchmarks with minimal computational overhead.

WARP: Wide Attention with Rich Projections for Image Super-Resolution

Addressing the inefficient multi-hop communication of deep shifted-window super-resolution transformers, WARP introduces a wide-attention architecture leveraging 2D RoPE and a cross-layer shared rich nonlinear projection, matching or outperforming deep 30-block SOTA models with only 12 blocks and 20M parameters.

Zero-Shot Inference-Time Rectification for Real-World Arbitrary-Scale Super-Resolution

The Continuous Degradation Rectifier (CDR) combines content purification, degradation-scale decoupling, and conditional diffusion to convert real-world low-resolution inputs into the bicubic-style images a frozen super-resolution model already masters; it brings CiaoSR a 4.6858 dB PSNR gain on RealSR ×2, though not all datasets and metrics improve in step.