Linear operators, probability, and ill-posed estimation
The null space of the forward operator determines what no reconstruction method can recover from the measurement alone.
Image restoration and inverse problems
Image restoration should begin with the measurement, not the architecture. Write a
forward model such as y = A(x) + ε: an unknown image x is transformed by an
operator A, then corrupted by noise ε, producing the observation y. Recovery
is difficult when information is destroyed, the noise is unknown, or many images
are compatible with the same measurement.
unknown image x ── forward operator A + noise ε ──→ measurement y ↑ │ └── reconstruction using measurement evidence + an explicit priorThe prior may be architectural, learned from paired data, imposed by cycle consistency, or represented by a generative model. It can prefer plausible images; only an explicit likelihood or data-consistency rule says how the reconstruction must agree with the observed measurement.
Separate four contracts
Section titled “Separate four contracts”| Contract | Required information | Central risk |
|---|---|---|
| supervised restoration | aligned clean/corrupted pairs and their degradation process | synthetic training degradation differs from deployment |
| paired translation | aligned source and target images | adversarial realism can compete with reconstruction fidelity |
| unpaired translation | domain samples plus structural assumptions | cycle consistency permits semantic changes |
| generative inverse reconstruction | learned prior, forward operator, likelihood, and sampler | plausible content is mistaken for measured content |
Evaluation must be plural
Section titled “Evaluation must be plural”No single metric establishes successful reconstruction. Report distortion metrics
such as PSNR when a reference is valid, structural or perceptual measures with their
limitations, forward consistency under A, uncertainty or multiple compatible
solutions, subgroup and degradation shift, and downstream task utility when that
task is part of the intended use.
Before comparing two methods, match training data, degradation simulator, input resolution, parameter and compute budget, model-selection access, and inference iterations. Then test blur, noise, resolution, compression, and domain mismatch one factor at a time.
The route below traces direct convolutional predictors, paired and unpaired translation, plug-and-play and untrained architectural priors, transformer restoration, diffusion foundations, and measurement-conditioned posterior samplers.
33 maintained papers · literature reviewed through 2 October 2026
An inverse problem begins with a forward measurement model such as y = A(x) + noise and asks which image x is supported by y. Super-resolution, denoising, deblurring, translation, and generative reconstruction use related architectures but make different assumptions about A, paired supervision, noise, and identifiability. A learned image prior can make outputs plausible; it does not by itself make them measurement-consistent or faithful. Every result should therefore separate distortion, perceptual quality, data consistency, uncertainty, and hallucination risk.
Prerequisite contract
Generative models and learned image priors
Diffusion and score models supply priors or samplers; inverse reconstruction additionally needs a likelihood or data-consistency rule.
Vision evaluation and robustness
PSNR, perceptual similarity, realism, and task utility can rank reconstructions differently.
How the field developed
2014–2017
Direct convolutional reconstruction and residual prediction
SRCNN learns an end-to-end mapping from an interpolated low-resolution image to high resolution. DnCNN predicts residual noise rather than the clean image, while EDSR removes batch normalization and scales a deeper residual network for super-resolution. Each method is trained for a specified synthetic degradation; success outside that degradation is a separate claim.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 01.01 Learning a Deep Convolutional Network for Image Super-Resolution SRCNN showed that single-image super-resolution could be learned as one fully convolutional mapping rather than a hand-separated patch pipeline. The low-resolution image is first enlarged by fixed bicubic interpolation; three learned convolutional stages then extract overlapping patch features, map them nonlinearly into high-resolution representations, and reconstruct the output. Training all filters jointly replaces dictionaries, iterative sparse coding, patch projection, and overlap averaging with a feed-forward network. — 01.02 Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising DnCNN reframed discriminative image denoising as residual prediction: a VGG-like fully convolutional network estimates the corruption in a noisy observation, and subtraction produces the clean estimate. Its practical contribution joins a deep receptive field, residual learning, and batch normalization so one feed-forward model can replace iterative denoisers. The paper also demonstrates that a single blind model can cover a continuum of Gaussian noise levels, while a separately defined multi-task model handles Gaussian denoising, several super-resolution scales, and JPEG deblocking. — 01.03 Enhanced Deep Residual Networks for Single Image Super-Resolution EDSR demonstrates that a high-level recognition ResNet should not be copied unchanged into low-level super-resolution. Removing batch normalization frees memory and avoids normalizing away useful image range; widening to 256 feature channels and deepening to 32 residual blocks supplies capacity; residual scaling by 0.1 stabilizes that large network; and L1 loss improves convergence. Its companion MDSR shares a deep trunk across x2, x3, and x4 while keeping scale-specific entry and upsampling modules. The plus suffix means eight-way test-time self-ensemble, not another trained architecture. — Work: Train matched direct and residual predictors under known blur, downsampling, and noise, then test operator and noise mismatch.
Exit evidence: Write the forward operator, training pairs, residual target, receptive field, and in- versus out-of-degradation metrics.
2017
Paired adversarial translation and unpaired cycle consistency
Pix2Pix combines a conditional GAN with a paired reconstruction term, requiring aligned source–target examples. CycleGAN replaces pairs with adversarial marginal matching plus a learned inverse and cycle-consistency losses. Cycle consistency constrains round trips; it does not prove that semantics, pathology, identity, or measurement evidence are preserved.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 02.01 Image-to-Image Translation with Conditional Adversarial Networks Pix2pix treats image-to-image translation as one paired conditional adversarial problem. A U-Net generator maps source image x to output G(x), preserving spatial detail through encoder-decoder skip connections. A PatchGAN discriminator judges overlapping local patches of the source-output pair as real or fake. Conditional adversarial loss encourages realistic local structure, while an L1 reconstruction term anchors the output to its aligned target and discourages arbitrary hallucination. Official implementation 02.02 Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks CycleGAN learns translation from unpaired image collections X and Y. Generator G maps X to Y and discriminator DY enforces target-domain appearance; generator F and discriminator DX learn the reverse. Cycle consistency requires F(G(x)) to reconstruct x and G(F(y)) to reconstruct y, constraining otherwise underdetermined adversarial mappings. Optional identity loss discourages unnecessary color changes when an image is already in the target domain. Official implementation Work: Compare paired and unpaired translation on the same domains with held-out correspondences and semantic controls.
Exit evidence: Separate adversarial realism, paired reconstruction, cycle consistency, identity constraints, and semantic error.
2013–2018
Denoiser modules and untrained architectural priors
Plug-and-Play Priors split a model-based objective with ADMM, preserve an explicit data-fidelity update, and replace the prior proximal step with a reusable denoiser—even when that denoiser has no known regularizer. Deep Image Prior instead optimizes the weights of a randomly initialized convolutional generator for one measured image; the network architecture and optimization trajectory bias the solution before any dataset training. PnP supplies a modular denoiser prior, whereas DIP supplies a per-instance architectural parameterization, so neither should be described as supervised restoration.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 03.01 Plug-and-Play Priors for Model Based Reconstruction Plug-and-Play Priors turn model-based inversion into two interchangeable modules: one update enforces the physical forward model and observed data, while another update injects image knowledge through a denoising operator. ADMM variable splitting is the interface. The decisive move is that the denoiser need not be written as the proximal map of a known regularizer, so K-SVD, BM3D, PLOW, total variation, or a discrete reconstruction routine can be inserted without redesigning the tomography solver. Author-hosted paper · Authors’ Plug-and-Play software and reading hub 03.02 Deep Image Prior Deep Image Prior (DIP) represents the unknown image as the output of a randomly initialized convolutional generator fθ(z), where z is fixed noise or another fixed input. For one corrupted observation, it optimizes only θ so that the known degradation of fθ(z) matches the observation. No external image dataset is used. The claim is not that an untrained network cannot fit noise; it is that convolutional architecture and gradient-descent trajectory fit natural image structure substantially earlier than unstructured noise, creating a useful optimization-time prior. Official project and supplementary material · Official implementation · arXiv record Work: Implement the same linear operator with an ADMM denoiser step and a DIP parameterization; hold the measurement, stopping rule, and metric code fixed.
Exit evidence: Identify the data-consistency update, denoiser or generator, optimization variable, regularization strength, stopping criterion, and convergence evidence.
2021–2022
Transformer context for high-resolution restoration
SwinIR adapts residual shifted-window Transformer blocks to super-resolution, denoising, and artifact reduction. Restormer redesigns attention and feed-forward computation for high-resolution restoration by mixing information efficiently across channels and space. Their context mechanisms differ, but both still inherit the degradation model and supervision used for training.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 04.01 SwinIR: Image Restoration Using Swin Transformer SwinIR turns the hierarchical shifted-window Transformer into a reusable image-restoration backbone. A 3-by-3 convolution first lifts a degraded image into shallow features; residual Swin Transformer blocks (RSTBs) then alternate local-window and shifted-window attention while retaining convolutional residual paths; a task-specific head reconstructs super-resolution, denoising, or JPEG-artifact-reduction output. Its central idea is not unrestricted global attention: it is efficient local attention whose shifted windows exchange information across boundaries, embedded inside a strong residual restoration pipeline. — 04.02 Restormer: Efficient Transformer for High-Resolution Image Restoration Restormer is a four-level encoder-decoder for high-resolution restoration whose key move is to transpose self-attention from an HW-by-HW spatial matrix to a C-by-C channel cross-covariance matrix. Multi-Dconv head transposed attention (MDTA) enriches Q, K, and V with pointwise and depthwise convolutions before channel attention; a gated depthwise-convolution feed-forward network (GDFN) filters the transformed features. The result has linear rather than quadratic dependence on spatial size for the attention map while retaining full-image evidence indirectly through channel statistics. — Work: Match parameter count and training data, then compare a convolutional baseline, window attention, and efficient restoration attention across image size.
Exit evidence: Report receptive context, tiling or padding policy, peak memory, throughput, boundary artifacts, and restoration metrics.
2020–2022
Diffusion, score, latent, and sampler design as learned priors
DDPM learns denoising transitions; score-SDE unifies score models in continuous time; DDIM changes the sampling trajectory without retraining the DDPM objective; latent diffusion moves denoising into a learned compressed space; and EDM clarifies noise parameterization, preconditioning, training weights, and numerical sampling. These papers establish generative priors and samplers. Turning them into inverse solvers requires an explicit measurement likelihood or data-consistency update tied to A.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 05.01 Denoising Diffusion Probabilistic Models DDPM fixes a Markov chain that gradually destroys data with small Gaussian-noise steps and learns Gaussian reverse transitions that denoise one step at a time. Its practical parameterization trains a time-conditioned U-Net to predict the exact noise used to create a randomly selected corrupted state, connecting variational diffusion training to multi-scale denoising score matching. — 05.02 Score-Based Generative Modeling through Stochastic Differential Equations Score-SDE replaces a finite noise ladder with a continuous stochastic differential equation that carries data to a tractable prior. Once a time-dependent network estimates every marginal score, the same model defines a reverse-time stochastic sampler, a deterministic probability-flow ODE with identical one-time marginals, predictor–corrector hybrids, likelihood evaluation, and conditional inverse-problem solvers. — 05.03 Denoising Diffusion Implicit Models DDIM shows that a pretrained DDPM noise predictor is compatible with a family of non-Markovian inference joints that retain the same noisy marginals and training objective. Setting reverse-step stochasticity to zero yields a deterministic implicit generator that can traverse a subsequence of timesteps, accelerating sampling and making the initial noise a reproducible latent code. — 05.04 High-Resolution Image Synthesis with Latent Diffusion Models Latent Diffusion moves the expensive denoising process from RGB pixels into a mildly compressed two-dimensional autoencoder latent, then decodes the final latent once. A reusable first stage removes perceptually less important detail; a U-Net learns the remaining latent distribution; cross-attention maps text, layouts, classes, or other tokens into intermediate denoising features. — 05.05 Elucidating the Design Space of Diffusion-Based Generative Models EDM reframes score-based diffusion as a modular engineering design space: noise parameterization, numerical solver, step schedule, stochasticity, network preconditioning, loss weighting, and training-noise distribution can be studied separately instead of being treated as one inseparable theoretical package. Code Work: Use one pretrained prior with several samplers, then add an explicit forward-operator likelihood and measure consistency separately from realism.
Exit evidence: Identify the prior, state variable, noise schedule, sampler, latent decoder, measurement term, and uncertainty interpretation.
2022–2023
Measurement-conditioned diffusion posterior sampling
DDRM derives efficient conditional reverse transitions for linear operators by working in the singular-vector basis of the measurement model; it reuses a pretrained diffusion prior and reports useful reconstructions in 20 neural-function evaluations, but requires a tractable linear degradation and its spectral decomposition. Diffusion Posterior Sampling adds an approximate measurement-likelihood gradient to the unconditional reverse score, using a denoised estimate of the clean image; this handles noisy nonlinear operators such as phase retrieval and non-uniform deblurring without a hard projection, but introduces an approximation and task-sensitive guidance step.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 06.01 Denoising Diffusion Restoration Models Denoising Diffusion Restoration Models (DDRM) reuse one pretrained denoising diffusion model to sample reconstructions for many known linear inverse problems without retraining for each degradation. The method rotates the linear measurement y=Hx+z into the singular-vector basis of H, where measured and unmeasured coordinates can receive different conditional reverse-diffusion updates. Its central promise is operator-time flexibility plus speed: the paper reports strong ImageNet restoration with 20 neural-function evaluations rather than hundreds or thousands. Official implementation · arXiv record 06.02 Diffusion Posterior Sampling for General Noisy Inverse Problems Diffusion Posterior Sampling (DPS) turns an unconditional score-based diffusion model into a solver for noisy linear and nonlinear inverse problems. At each reverse step it estimates the clean image x̂0 from the current noisy state, evaluates how the differentiable forward operator A(x̂0) disagrees with measurement y, and adds the gradient of that measurement likelihood to the learned prior score. Unlike projection-based noiseless solvers, DPS deliberately avoids a hard measurement-subspace projection, allowing Gaussian or Poisson noise and nonlinear operators such as Fourier phase retrieval and non-uniform deblurring. Official implementation · arXiv record Work: Apply DDRM and DPS to matched linear Gaussian measurements, then add one nonlinear operator for DPS; audit operator access, likelihood, NFE count, and posterior diversity.
Exit evidence: Explain why DDRM needs linear spectral structure, how DPS differentiates through the clean-image estimate, and which metric measures realism versus fidelity.
2022–2026
Simple backbones and continual degradation restoration
NAFNet shows that a carefully designed multiplicative restoration block can outperform more ornate baselines without standard nonlinear activations; Restoring Without Forgetting isolates low-rank degradation adapters and routes among them over a stream.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 07.01 Simple Baselines for Image RestorationPublished in the ECCV 2022 proceedings; ECVA hosts the canonical open proceedings PDF. Core mechanism: nonlinear activations removed from a restoration baseline. NAFNet starts from a plain encoder-decoder restoration baseline and removes conventional ReLU/GELU and sigmoid activations, replacing key behavior with channel splitting and elementwise multiplication plus simplified attention-like channel scaling. This reading boundary matters: the network is free of conventional activation functions but remains nonlinear through multiplication, and its reported strength concerns supervised denoising/deblurring protocols rather than every inverse problem. Official implementation and models 07.02 Restoring Without Forgetting: Continual Learning Across Image DegradationsarXiv v1 preprint submitted 24 August 2026; validation covers the reported five degradation domains rather than arbitrary composite corruption. Core mechanism: frozen restoration backbone with isolated low-rank adapters. The method freezes a shared restoration backbone, learns a separate low-rank adapter for each arriving degradation, stores feature prototypes for degradation recognition, and routes an input to the corresponding adapter so earlier domains are not overwritten. This reading boundary matters: the architecture avoids parameter overwrite across the five studied degradation domains, not automatic handling of unseen, mixed, continuously varying, or misrouted corruptions. Official implementation Work: Reproduce NAFNet and degradation-specific low-rank adapters with identical corruptions, train/validation splits, parameter and update budgets; then stream degradations in several orders with and without replay or routing.
Exit evidence: Report restoration quality, perceptual and pixel metrics, parameters, MACs, measured latency, adapter memory, order variance, forgetting, forward transfer, and router errors.
2026
Codec-audited SAR-to-EO flow matching
ReFlowSET makes latent-codec selection part of the experiment, then trains a compact conditional DiT with SAR conditioning and clean-EO representation guidance.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 08.01 ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image TranslationarXiv v1 submitted 1 September 2026. The exact title follows the current arXiv record, superseding an earlier feed title. Code, weights, and an official project page were available; no peer-reviewed record was verified. Core idea: audit the latent codec on both radar conditions and optical targets before training the generator, then learn conditional rectified flow in the selected latent space. ReFlowSET uses a from-scratch conditional DiT with separate SAR and noisy-EO streams, delayed channel fusion, and a frozen vision-foundation-model alignment loss that guides training but is removed at inference. Official implementation and checkpoints · Official project page Work: Audit round-trip reconstruction for both modalities before training the translation model, then compare pixel, perceptual, distributional, and task metrics.
Exit evidence: Report codec, paired-data assumptions, cloud/season shift, training-only teacher cost, and geospatial error cases.
2019
Recognition-guided cross-resolution face hallucination
RIM reconstructs a higher-resolution face with a prior-aided hallucination network while an assistant learns the teacher–student feature residual for recognition.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 09.01 Cross-Resolution Face Recognition via Prior-Aided Face Hallucination and Residual Knowledge DistillationarXiv technical report submitted 26 May 2019; no peer-reviewed version or released implementation was located by 2026-09-02. RIM first hallucinates a high-resolution face from a 28 by 28 input, then recognizes it with a compact network whose assistant predicts blockwise feature residuals relative to a high-resolution teacher. The residual sum is the transferable representation; super-resolution and recognition are trained together rather than treated as independent modules. — Work: Measure pixel and perceptual reconstruction, identity verification, hallucinated detail, assistant cost, and low-resolution-only deployment separately.
Exit evidence: Test native low-resolution and domain-shifted faces, with uncertainty and identity-preserving failure analysis.
2025–2026
Optical observability, phase recovery and calibrated deblurring
Compare three ways of making an inverse problem identifiable: expose latent focus texture through bandwidth/exposure control, measure phase differences on a short-path graph, or capture complementary focal planes and fit spatially varying blur. In each case, the sensing operator matters before the reconstruction prior.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 10.01 A second-order theory of texture for depth from focusAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Author spelling follows the conference PDF. The official project still lists code and data as forthcoming. A surface can be textureless in its average reflectance yet have measurable image texture from unresolved microscopic roughness. This second-order texture is subjective speckle: intensity fluctuations around the ensemble mean. Narrowband filtering reduces spectral averaging and can make that texture strong enough for passive depth from focus, provided exposure and illumination yield sufficient signal-to-noise ratio. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record 10.02 Provable and Robust Wavefront Sensing via Self-Reference InterferometryAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Official implementation inspected, not executed; its README labels the sequential v2 least-squares tail unfinished and uses the vectorized v3 path. Coprime self-reference interferometry converts phase retrieval into propagation of measured phase differences on a connected pixel graph. Interfering a field with shifted copies of itself removes the need for a separate known reference beam. Carefully selected coprime shifts shorten propagation paths, reducing accumulated noise; equal-hop estimates are averaged and optional global least squares can refine the result. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed) 10.03 Broadband Full Field of View Imaging with Computational MirrorsAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the conference-hosted PDF; the official program/arXiv also use “Broadband Wide Field of View Imaging with Computational Mirrors”. Code/data are announced but not yet linked as released. Computational Mirrors turns achromatic but spatially blurred reflective optics into a usable broadband camera by combining a sparse focal stack with SeidelConv. This calibrated forward model warps, blurs and spatially weights several image branches to represent off-axis aberrations; multi-image plug-and-play reconstruction then recovers a sharp image. Optical measurement design and inversion are the contribution, not a new image classifier. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record Work: Which measurement creates the missing evidence, and which ambiguity remains after reconstruction?
Exit evidence: Do not turn a recoverability proxy into ground-truth depth accuracy, relative phase into absolute phase, or multi-exposure broadband capture into snapshot video. Calibration, light throughput, scene motion and operator assumptions remain part of the method.
2025–2026
Physics-guided residual refinement of frozen geometry models
Poppy uses test-time image offsets, delayed normal offsets and radiance decomposition to align a frozen RGB geometric prior with polarimetric measurements. This offers a concrete residual-correction mechanism to contrast with teacher-supervised residual KD, while preserving their different supervision and cost.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 11.01 Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal EstimationAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the PDF; the program calls the task “Surface Normal Estimation”. Training-free means no backbone retraining, not no test-time optimization. Poppy corrects a frozen RGB normal estimator with polarization measurements at test time. It learns image offsets for scene-wide corrections, normal offsets for fine detail, and a specular-radiance map that explains mixed reflection. This is physics-guided residual optimization around a pretrained model, not teacher–student knowledge distillation. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed) Work: Is the residual supervised by another model, a physical observation or task labels, and what actually changes at test time?
Exit evidence: Training-free backbone adaptation still incurs repeated backpropagation and needs polarization measurements; orthographic assumptions and sensor noise can reverse local-refinement gains.
2025–2026
Physics-informed optical fields and focus-stack geometry
Event-domain refractive inversion estimates evolving air density, while calibrated focus-stack attention estimates scene depth. Both exploit a known optical observation process; compare acquisition metadata, scene-specific optimization and synthetic-to-real validation before treating them as generic image encoders.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 12.01 Event-based Sparse-view Background-Oriented Schlieren TomographyECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #377. Publisher proceedings/version of record was not verified on 2026-09-12. EventBOS reconstructs a continuous four-dimensional air-density field directly from event streams. A coordinate MLP predicts density, velocity and pressure; differentiable refractive ray tracing turns density into changes in a textured background, which are matched to measured events. Fluid equations regularize the otherwise ambiguous one- or two-view inverse problem. ECCV Spotlight program · Conference manuscript 12.02 Zero-shot Depth from DefocusECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12. FOSSA learns metric depth from a calibrated focus stack using attention across focus settings. It alternates ordinary image-token processing with per-location stack attention conditioned on focus distance, then pools the stack and predicts dense depth. The accompanying ZEDD benchmark tests real-world zero-shot transfer rather than within-dataset fitting. ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link) Work: Which unknown is identifiable from the actual measurements, and what boundary, calibration or scene-static assumptions supply missing information?
Exit evidence: High event acquisition rate is not real-time optimization; multiple focus images provide different evidence from one monocular image.
2025–2026
Physical observability and task-aware polarization reconstruction
GenPolar predicts Stokes components from intensity using observability-aware supervision and one-step diffusion distillation. PolarAPP instead starts from measured polarization mosaics and couples demosaicking with downstream objectives through a learned metric. Keep generative priors separate from sensor-acquired polarization.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 13.01 Stokes-Informed Diffusion for Robust Linear Polarization EstimationECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #5749. Publisher proceedings/version of record was not verified on 2026-09-12. GenPolar predicts RGB-channel linear Stokes components from an intensity image, then derives polarization magnitude and angle analytically. It combines observability-aware physical supervision with multi-step-to-one-step diffusion distillation and VAE-encoder LoRA. This targets stable polarimetric cues rather than merely plausible-looking angle maps. ECCV Spotlight program · Conference manuscript 13.02 PolarAPP: Beyond Polarization Demosaicking for Polarimetric ApplicationsECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #5713. Publisher proceedings/version of record was not verified on 2026-09-12. PolarAPP jointly optimizes polarization demosaicking and a downstream task using a meta-learned feature-alignment metric. An equivalent-imaging self-consistency constraint improves reconstruction without treating rearranged mosaic pixels as perfect targets. Training-time alignment modules disappear at inference, leaving a demosaicker, analytic polarization cues and the task network. ECCV Spotlight program · Conference manuscript Work: Is the supervision a physically observable target, a synthetic reconstruction target, a downstream task signal or a teacher distribution?
Exit evidence: Weak-polarization angle masks change the evaluated population. Task-aware meta-learning is not automatically teacher–student KD, and metric improvements need not agree.
2025–2026
Resolution transfer and calibrated geophysical fields
BaguanHR creates additional fine-resolution atmospheric training fields through super-resolution, whereas Sparse2Cloud predicts dense cloud geometry from sparse active-sensor tracks and calibrates quantile intervals. Compare transferred information, geographic or temporal splits and uncertainty, not just grid density.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 14.01 Pushing the Limits of High-Resolution Weather Forecasting through Data ScalingECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #7258. Publisher proceedings/version of record was not verified on 2026-09-12. BaguanHR transfers high-resolution data rather than merely transferring a coarse forecast model. Variable-specific super-resolution expands scarce 0.1-degree atmospheric training data from long 0.25-degree ERA5 records, then a high-resolution forecaster learns from synthetic-plus-real fields. Adaptive rollout weighting and replay target long-horizon error accumulation. ECCV Spotlight program · Conference manuscript · Author project (paper/program link) 14.02 Ice Cloud Geometry Retrieval with Calibrated Uncertainty from Passive Satellite ImageryECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #12985. Publisher proceedings/version of record was not verified on 2026-09-12. Selected aggregate values follow Table 1. Table 2 reports a different G-Q aggregate R2 (0.725 vs 0.731); keep table-specific provenance. CQR guarantees marginal coverage only under exchangeability, not arbitrary subgroups or global domain shift. Sparse2Cloud learns dense ice-cloud geometry from narrow satellite supervision tracks and attaches conformal prediction intervals. A ConvNeXt U-Net predicts eight cloud targets and three quantiles per target, conditioned on passive VIIRS channels and ERA5 profiles. A five-model quantile ensemble plus calibration separates point accuracy from interval coverage. ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link) Work: Does a finer output represent measured extra information, learned prior interpolation or a calibrated uncertainty statement?
Exit evidence: Synthetic resolution does not guarantee missing fine physics. Marginal conformal coverage assumes exchangeability and is not a universal conditional guarantee under climate or geographic shift.
2025–2026
Implicit-field tokenization of biomedical graphs
VesselTok compresses pre-extracted centerline graphs through pseudo-radius occupancy reconstruction, then uses latent generative models for synthesis and repair. This connects shape representation, VAE bottlenecks and graph topology without claiming to infer vessels directly from raw CT.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 15.01 VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and GenerationECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10927. Publisher proceedings/version of record was not verified on 2026-09-12. VesselTok compresses complex biomedical centerline graphs into continuous latent tokens by reconstructing a pseudo-radius occupancy field. Centerline-conditioned attention preserves thin branching structure more efficiently than densely sampling its surface. The same latent supports graph generation and conditional repair, but it does not directly encode true vessel thickness. ECCV Spotlight program · Conference manuscript · Official implementation Work: What geometric and topological information survives token compression, and which attributes were never represented?
Exit evidence: Pseudo radius is not measured vessel thickness; high overlap can coexist with worse cycle topology, and token capacity can merge or omit small structures.
2025–2026
Second-order attention and lightweight image denoising
NLformer augments non-local first-order patch averaging with low-rank intra-patch statistics and a Woodbury-inspired inverse update. The resulting attention and feed-forward blocks offer a concrete example of turning a classical estimator into a trainable architecture.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 16.01 Bayesian Self-Attention with Intra-Token Modeling for Lightweight Denoising TransformersECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #7178. Publisher proceedings/version of record was not verified on 2026-09-12. Conference program and code use 'Local Pixel Correlations'; checked conference PDF uses 'Intra-Token Modeling'. Both titles describe submission 7178; PDF title is canonical here. NLformer treats denoising attention as a combination of non-local patch averaging and a local second-order correction. Its dual attention branches model first-order inter-token relations and low-rank intra-token statistics; a Woodbury-based feed-forward module makes the associated matrix inversion affordable. The result is a compact denoiser, not a distilled student. ECCV Spotlight program · Conference manuscript · Official implementation Work: Which part estimates a mean, which part corrects local covariance, and how is the matrix inverse made tractable?
Exit evidence: A simplified Bayesian derivation is not an exact guarantee for real camera noise. Fewer parameters need not mean fewer FLOPs than every lightweight competitor.
2025–2026
Illuminant coordinates and camera color calibration
CCM-MLP conditions a compact color matrix on two-dimensional illuminant chromaticity, relaxing the traditional one-dimensional color-temperature assumption. Read it as image formation and camera calibration: the key change is the conditioning space, not a generic larger network.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 17.01 Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera ColorECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #1413. Publisher proceedings/version of record was not verified on 2026-09-12. CCM-MLP replaces one-dimensional color-temperature interpolation with a two-dimensional illuminant-chromaticity-conditioned color correction matrix. A tiny MLP maps a white point to eight free matrix entries, enabling an ordinary camera pipeline to handle LED illuminants away from the Planckian locus. This is calibrated color mapping, not a new white-balance estimator. ECCV Spotlight program · Conference manuscript · Author project (paper/program link) Work: How does off-Planckian illumination expose the limits of temperature-only interpolation, and who estimates the white point?
Exit evidence: White-point estimation error and metamerism remain. Angular-error gains can differ from perceptual color-error rankings and do not establish stain-normalization results.
Research branches
Degradation-specific supervised restoration
Is the forward operator known, simulated accurately, and represented in the training pairs?
Learning a Deep Convolutional Network for Image Super-Resolution · Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising · Enhanced Deep Residual Networks for Single Image Super-Resolution
Paired and unpaired image translation
Which correspondence or structural constraint prevents a realistic but incorrect translation?
Image-to-Image Translation with Conditional Adversarial Networks · Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks
High-resolution transformer restoration
How is long-range context obtained without prohibitive memory or boundary artifacts?
SwinIR: Image Restoration Using Swin Transformer · Restormer: Efficient Transformer for High-Resolution Image Restoration
Modular and per-instance priors
Is prior information supplied by an external denoiser or by the generator architecture and stopping time?
Plug-and-Play Priors for Model Based Reconstruction · Deep Image Prior
Diffusion foundations and inverse solvers
How are the learned prior and measurement likelihood coupled during inference?
Denoising Diffusion Probabilistic Models · Score-Based Generative Modeling through Stochastic Differential Equations · Denoising Diffusion Implicit Models · High-Resolution Image Synthesis with Latent Diffusion Models · Elucidating the Design Space of Diffusion-Based Generative Models · Denoising Diffusion Restoration Models · Diffusion Posterior Sampling for General Noisy Inverse Problems
Simple restoration backbones and continual degradation adapters
Does the gain come from a simpler multiplicative block, retained task-specific capacity, correct degradation routing, or access to prior data?
Simple Baselines for Image Restoration · Restoring Without Forgetting: Continual Learning Across Image Degradations
Cross-sensor inverse translation
Does the latent representation preserve what the inverse problem must reconstruct?
ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation
Cross-resolution restoration for recognition
Does reconstructed detail recover supported identity evidence or invent a recognition-friendly face?
Optical observability, phase recovery and calibrated deblurring
Which measurement creates the missing evidence, and which ambiguity remains after reconstruction?
A second-order theory of texture for depth from focus · Provable and Robust Wavefront Sensing via Self-Reference Interferometry · Broadband Full Field of View Imaging with Computational Mirrors
Physics-guided residual refinement of frozen geometry models
Is the residual supervised by another model, a physical observation or task labels, and what actually changes at test time?
Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Physics-informed optical fields and focus-stack geometry
Which unknown is identifiable from the actual measurements, and what boundary, calibration or scene-static assumptions supply missing information?
Event-based Sparse-view Background-Oriented Schlieren Tomography · Zero-shot Depth from Defocus
Physical observability and task-aware polarization reconstruction
Is the supervision a physically observable target, a synthetic reconstruction target, a downstream task signal or a teacher distribution?
Stokes-Informed Diffusion for Robust Linear Polarization Estimation · PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications
Resolution transfer and calibrated geophysical fields
Does a finer output represent measured extra information, learned prior interpolation or a calibrated uncertainty statement?
Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling · Ice Cloud Geometry Retrieval with Calibrated Uncertainty from Passive Satellite Imagery
Implicit-field tokenization of biomedical graphs
What geometric and topological information survives token compression, and which attributes were never represented?
Second-order attention and lightweight image denoising
Which part estimates a mean, which part corrects local covariance, and how is the matrix inverse made tractable?
Bayesian Self-Attention with Intra-Token Modeling for Lightweight Denoising Transformers
Illuminant coordinates and camera color calibration
How does off-Planckian illumination expose the limits of temperature-only interpolation, and who estimates the white point?
Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera Color
Comparison table
| Family or stage | Scientific distinction | Main risk | Anchors |
|---|---|---|---|
| Direct / residual reconstruction | Learns a deterministic mapping for an explicit paired degradation process. | Synthetic operator mismatch can dominate architecture differences. | 2014 · 2017 · 2017 |
| Adversarial translation | Matches a target appearance distribution with paired or cycle constraints. | Realism objectives can alter unsupported content. | 2017 · 2017 |
| Transformer restoration | Expands contextual interaction while preserving image-resolution outputs. | Memory, tiling, and training degradation remain limiting factors. | 2021 · 2022 |
| Plug-in / architectural prior | Alternates an explicit forward-model step with a denoiser, or optimizes one untrained generator against the measurement. | A black-box denoiser may not correspond to an objective; DIP can eventually fit noise and depends on early stopping. | 2013 · 2018 |
| Diffusion / score prior | Represents a distribution and supplies the reverse process used by conditional inverse samplers. | A generic prior is not yet an inverse solver, and plausible samples may hide posterior ambiguity. | 2020 · 2021 · 2021 · 2022 · 2022 |
| Diffusion posterior solver | Conditions reverse diffusion with linear spectral transitions or an approximate likelihood gradient. | Operator restrictions, guidance approximations, NFE cost, and prior mismatch can alter fidelity. | 2022 · 2023 |
| Simple backbones and continual degradation restoration | NAFNet removes standard nonlinear activations in a carefully designed restoration block; continual restoration isolates low-rank degradation adapters and routes among them over a stream. | A favorable benchmark recipe can be attributed to “activation-free” design, while adapter growth, oracle task identity, or replay can hide continual-learning cost. | 2022 · 2026 |
| Codec-audited SAR-to-EO flow matching | Treats codec suitability and representation alignment as first-class conditional-generation choices. | Natural-image codecs and automated perceptual metrics can hide sensor-specific or geographic errors. | 2026 |
| Recognition-guided cross-resolution face hallucination | Couples super-resolution with an explicit residual feature assistant rather than optimizing perceptual quality alone. | Face-specific priors can hallucinate identity-bearing structure and do not directly justify pathology super-resolution. | 2019 |
| Optical observability, phase recovery and calibrated deblurring | Compare three ways of making an inverse problem identifiable: expose latent focus texture through bandwidth/exposure control, measure phase differences on a short-path graph, or capture complementary focal planes and fit spatially varying blur. In each case, the sensing operator matters before the reconstruction prior. | Do not turn a recoverability proxy into ground-truth depth accuracy, relative phase into absolute phase, or multi-exposure broadband capture into snapshot video. Calibration, light throughput, scene motion and operator assumptions remain part of the method. | 2026 · 2026 · 2026 |
| Physics-guided residual refinement of frozen geometry models | Poppy uses test-time image offsets, delayed normal offsets and radiance decomposition to align a frozen RGB geometric prior with polarimetric measurements. This offers a concrete residual-correction mechanism to contrast with teacher-supervised residual KD, while preserving their different supervision and cost. | Training-free backbone adaptation still incurs repeated backpropagation and needs polarization measurements; orthographic assumptions and sensor noise can reverse local-refinement gains. | 2026 |
| Physics-informed optical fields and focus-stack geometry | Event-domain refractive inversion estimates evolving air density, while calibrated focus-stack attention estimates scene depth. Both exploit a known optical observation process; compare acquisition metadata, scene-specific optimization and synthetic-to-real validation before treating them as generic image encoders. | High event acquisition rate is not real-time optimization; multiple focus images provide different evidence from one monocular image. | 2026 · 2026 |
| Physical observability and task-aware polarization reconstruction | GenPolar predicts Stokes components from intensity using observability-aware supervision and one-step diffusion distillation. PolarAPP instead starts from measured polarization mosaics and couples demosaicking with downstream objectives through a learned metric. Keep generative priors separate from sensor-acquired polarization. | Weak-polarization angle masks change the evaluated population. Task-aware meta-learning is not automatically teacher–student KD, and metric improvements need not agree. | 2026 · 2026 |
| Resolution transfer and calibrated geophysical fields | BaguanHR creates additional fine-resolution atmospheric training fields through super-resolution, whereas Sparse2Cloud predicts dense cloud geometry from sparse active-sensor tracks and calibrates quantile intervals. Compare transferred information, geographic or temporal splits and uncertainty, not just grid density. | Synthetic resolution does not guarantee missing fine physics. Marginal conformal coverage assumes exchangeability and is not a universal conditional guarantee under climate or geographic shift. | 2026 · 2026 |
| Implicit-field tokenization of biomedical graphs | VesselTok compresses pre-extracted centerline graphs through pseudo-radius occupancy reconstruction, then uses latent generative models for synthesis and repair. This connects shape representation, VAE bottlenecks and graph topology without claiming to infer vessels directly from raw CT. | Pseudo radius is not measured vessel thickness; high overlap can coexist with worse cycle topology, and token capacity can merge or omit small structures. | 2026 |
| Second-order attention and lightweight image denoising | NLformer augments non-local first-order patch averaging with low-rank intra-patch statistics and a Woodbury-inspired inverse update. The resulting attention and feed-forward blocks offer a concrete example of turning a classical estimator into a trainable architecture. | A simplified Bayesian derivation is not an exact guarantee for real camera noise. Fewer parameters need not mean fewer FLOPs than every lightweight competitor. | 2026 |
| Illuminant coordinates and camera color calibration | CCM-MLP conditions a compact color matrix on two-dimensional illuminant chromaticity, relaxing the traditional one-dimensional color-temperature assumption. Read it as image formation and camera calibration: the key change is the conditioning space, not a generic larger network. | White-point estimation error and metamerism remain. Angular-error gains can differ from perceptual color-error rankings and do not establish stain-normalization results. | 2026 |
Concept lineage
Learning a Deep Convolutional Network for Image Super-ResolutionEnhanced Deep Residual Networks for Single Image Super-Resolution
A shallow direct super-resolution mapping develops into a deeper residual backbone specialized for high-fidelity reconstruction.
Image-to-Image Translation with Conditional Adversarial NetworksUnpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks
Aligned pairs are replaced by adversarial marginal matching and cycle constraints, changing the identifiability contract.
SwinIR: Image Restoration Using Swin TransformerRestormer: Efficient Transformer for High-Resolution Image Restoration
Windowed spatial context is complemented by an attention design tailored to efficient high-resolution restoration.
Plug-and-Play Priors for Model Based ReconstructionDeep Image Prior
A reusable denoising operator inside explicit inversion is complemented by an image-specific neural parameterization whose architecture itself supplies the prior.
Denoising Diffusion Probabilistic ModelsScore-Based Generative Modeling through Stochastic Differential Equations
Discrete denoising transitions are connected to continuous-time score dynamics.
Denoising Diffusion Probabilistic ModelsDenoising Diffusion Implicit Models
The learned denoising model is retained while the reverse sampling path is changed.
Denoising Diffusion Probabilistic ModelsHigh-Resolution Image Synthesis with Latent Diffusion Models
Pixel-space denoising moves into a learned latent space to trade reconstruction detail for computational scale.
Denoising Diffusion Probabilistic ModelsDenoising Diffusion Restoration Models
An unconditional diffusion prior is converted into an efficient conditional sampler for linear measurements through operator-aware reverse transitions.
Score-Based Generative Modeling through Stochastic Differential EquationsDiffusion Posterior Sampling for General Noisy Inverse Problems
The unconditional reverse score is augmented with an approximate measurement-likelihood gradient so noisy nonlinear operators can guide sampling.
Failure modes to audit
- Omitting the forward operator, noise process, or boundary conditions from the task definition.
- Training on one synthetic degradation and claiming general restoration without mismatch tests.
- Treating perceptual realism or adversarial success as evidence of measurement fidelity.
- Reporting only PSNR, only a perceptual metric, or only a task score.
- Calling an unconditional generative prior an inverse solver without a likelihood or data-consistency mechanism.
- Treating a plug-and-play fixed point as the minimizer of a known regularizer without checking the denoiser assumptions.
- Optimizing Deep Image Prior to convergence without monitoring when structured signal gives way to fitted noise.
- Showing one reconstruction when the posterior is multimodal or the problem is non-identifiable.
Questions for every paper
- What is the forward operator A, its null space, and its noise model?
- Which information comes from the measurement and which comes only from the learned prior?
- Are training pairs aligned, synthetic, unpaired, or generated by a simulator?
- How are distortion, perceptual quality, consistency, uncertainty, and downstream utility measured separately?
- Which operator, noise, resolution, and domain shifts challenge the method?
- Does the method require an SVD, differentiable forward operator, explicit regularizer, paired training data, or early stopping?
Complete maintained paper route
Section titled “Complete maintained paper route”33 canonical readings · 2013–2026· route reviewed through October 2, 2026.
This is the complete structured list for this chapter, ordered by an explicit prerequisite-aware teaching sequence. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.
Show all 33 readings in maintained order
Plug-and-Play Priors for Model Based Reconstruction
Singanallur V. Venkatakrishnan, Charles A. Bouman, and Brendt Wohlberg
Official artifacts: Author-hosted paper · Authors’ Plug-and-Play software and reading hub
Metadata checked
Broadband Full Field of View Imaging with Computational Mirrors
Vishwanath Saragadam, Niki Nezakati, Amit Roy-Chowdhury, and Vivek Boominathan
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the conference-hosted PDF; the official program/arXiv also use “Broadband Wide Field of View Imaging with Computational Mirrors”. Code/data are announced but not yet linked as released.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Metadata checked
Provable and Robust Wavefront Sensing via Self-Reference Interferometry
Nebiyou Yismaw, Vishwanath Saragadam, Aswin C. Sankaranarayanan, and M. Salman Asif
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Official implementation inspected, not executed; its README labels the sequential v2 least-squares tail unfinished and uses the vectorized v3 path.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed)
Metadata checked
A second-order theory of texture for depth from focus
Sreekar Sai Ranganathan and Ioannis Gkioulekas
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Author spelling follows the conference PDF. The official project still lists code and data as forthcoming.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Metadata checked
Event-based Sparse-view Background-Oriented Schlieren Tomography
Xinyu Zhou, Shihao Hu, Peiqi Duan, Chao Xu, and Boxin Shi
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #377. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript
Metadata checked
Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera Color
SaiKiran Tedla, Joshua Little, Hakki C. Karaimer, and Michael S. Brown
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #1413. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Author project (paper/program link)
Metadata checked
Learning a Deep Convolutional Network for Image Super-Resolution
Chao Dong et al.
Metadata checked
Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising
Kai Zhang et al.
Metadata checked
Enhanced Deep Residual Networks for Single Image Super-Resolution
Bee Lim et al.
Metadata checked
Image-to-Image Translation with Conditional Adversarial Networks
Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros
Official artifacts: Official implementation
Metadata checked
Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks
Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros
Official artifacts: Official implementation
Metadata checked
Deep Image Prior
Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky
Official artifacts: Official project and supplementary material · Official implementation · arXiv record
Metadata checked
Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Irene Kim, Sai Tanmay Reddy Chakkera, Alexandros Graikos, Dimitris Samaras, and Akshat Dave
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the PDF; the program calls the task “Surface Normal Estimation”. Training-free means no backbone retraining, not no test-time optimization.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed)
Metadata checked
Zero-shot Depth from Defocus
Yiming Zuo, Hongyu Wen, Venkat Subramanian, Patrick Chen, Karhan Kayan, Mario Bijelic, Felix Heide, and Jia Deng
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)
Metadata checked
VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation
Chinmay Prabhakar, Bastian Wittmann, Tamaz Amiranashvili, Paul Büschl, Ezequiel De la Rosa, Julian McGinnis, Benedikt Wiestler, Bjoern Menze, and Suprosanna Shit
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10927. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation
Metadata checked
Cross-Resolution Face Recognition via Prior-Aided Face Hallucination and Residual Knowledge Distillation
Hanyang Kong, Jian Zhao, Xiaoguang Tu, Junliang Xing, Shengmei Shen, and Jiashi Feng
arXiv technical report submitted 26 May 2019; no peer-reviewed version or released implementation was located by 2026-09-02.
Metadata checked
Denoising Diffusion Probabilistic Models
Jonathan Ho, Ajay Jain, and Pieter Abbeel
Metadata checked
Score-Based Generative Modeling through Stochastic Differential Equations
Yang Song et al.
Metadata checked
Denoising Diffusion Implicit Models
Jiaming Song, Chenlin Meng, and Stefano Ermon
Metadata checked
SwinIR: Image Restoration Using Swin Transformer
Jingyun Liang et al.
Metadata checked
Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling
Yang Zhao, Peisong Niu, Tian Zhou, Ziqing Ma, Guanlong Ma, Rong Jin, Huiling Yuan, and Liang Sun
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #7258. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Author project (paper/program link)
Metadata checked
Ice Cloud Geometry Retrieval with Calibrated Uncertainty from Passive Satellite Imagery
Ayush Prasad
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #12985. Publisher proceedings/version of record was not verified on 2026-09-12. Selected aggregate values follow Table 1. Table 2 reports a different G-Q aggregate R2 (0.725 vs 0.731); keep table-specific provenance. CQR guarantees marginal coverage only under exchangeability, not arbitrary subgroups or global domain shift.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)
Metadata checked
Restormer: Efficient Transformer for High-Resolution Image Restoration
Syed Waqas Zamir et al.
Metadata checked
Bayesian Self-Attention with Intra-Token Modeling for Lightweight Denoising Transformers
Runyang He, Zuowei Shen, and Hui Ji
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #7178. Publisher proceedings/version of record was not verified on 2026-09-12. Conference program and code use 'Local Pixel Correlations'; checked conference PDF uses 'Intra-Token Modeling'. Both titles describe submission 7178; PDF title is canonical here.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation
Metadata checked
Simple Baselines for Image Restoration
Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun
Published in the ECCV 2022 proceedings; ECVA hosts the canonical open proceedings PDF.
Official artifacts: Official implementation and models
Metadata checked
High-Resolution Image Synthesis with Latent Diffusion Models
Robin Rombach et al.
Metadata checked
Elucidating the Design Space of Diffusion-Based Generative Models
Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine
Official artifacts: Code
Metadata checked
Denoising Diffusion Restoration Models
Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song
Official artifacts: Official implementation · arXiv record
Metadata checked
Diffusion Posterior Sampling for General Noisy Inverse Problems
Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye
Official artifacts: Official implementation · arXiv record
Metadata checked
Stokes-Informed Diffusion for Robust Linear Polarization Estimation
Yidong Luo, Chenggong Li, Yuchao Feng, Boxin Shi, Junchao Zhang, and Xin Yuan
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #5749. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript
Metadata checked
PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications
Yidong Luo, Chenggong Li, Yunfeng Song, Ping Wang, Boxin Shi, Junchao Zhang, and Xin Yuan
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #5713. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript
Metadata checked
Restoring Without Forgetting: Continual Learning Across Image Degradations
Alif Ashrafee and Bartosz Krawczyk
arXiv v1 preprint submitted 24 August 2026; validation covers the reported five degradation domains rather than arbitrary composite corruption.
Official artifacts: Official implementation
Metadata checked
ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation
Jeonghyeok Do, Seungchul Lee, and Munchurl Kim
arXiv v1 submitted 1 September 2026. The exact title follows the current arXiv record, superseding an earlier feed title. Code, weights, and an official project page were available; no peer-reviewed record was verified.
Official artifacts: Official implementation and checkpoints · Official project page
Metadata checked