🖼️ Image Restoration¶

🧪 ICML2026 · 21 paper notes

📌 Same area in other venues: 📷 CVPR2026 (135) · 🔬 ICLR2026 (61) · 🤖 AAAI2026 (10) · 🧠 NeurIPS2025 (26) · 📹 ICCV2025 (31) · 🧪 ICML2025 (5)

🔥 Top topics: Diffusion Models ×11 · Super-Resolution ×5 · Image Restoration ×2

AnyMod-LLVE: Low-Light Video Enhancement with Modality-Agnostic Inference: Addressing the issue where "multi-modal low-light video enhancement collapses when event streams or infrared auxiliary modalities are unavailable during inference," AMNet utilizes a Spatial-Spectral Dual-Gated (S2DG) Translator to generate implicit representations of auxiliary modalities from degraded low-light RGB inputs. Combined with large-scale synthetic multi-modal pre-training, this allows stable enhancement regardless of modality availability during testing—achieving SOTA with RGB-only inference, with further gains when auxiliary modalities are provided.
Coevolutionary Continuous Discrete Diffusion: Make Your Diffusion Language Model a Latent Reasoner: This paper systematically compares continuous diffusion, discrete masked diffusion, and looped transformers across the dimensions of expressivity and trainability. It proves that "continuous diffusion" is strictly more expressive than discrete diffusion and can simulate looped transformers, but its practical performance is limited by decoding and representation space. Consequently, the paper proposes CCDD (Coevolutionary Continuous Discrete Diffusion)—diffusion performed simultaneously on the discrete token space and the contextual embedding space of a pre-trained LLM, with a single model for joint denoising. CCDD reduces perplexity by 25-35% compared to MDLM on LM1B/OWT and outperforms MDLM with 256 steps using only 8 sampling steps.
Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution: ASASR achieves the optimal balance between perceptual quality and structural fidelity in super-resolution by replacing the Flow Matching noise prior from isotropic Gaussian to Sobolev spectral coloring noise, combined with adversarial manifold guidance to generate hard negative samples, constructing the AS-DPO framework.
Consistent Diffusion Language Models: This paper points out that discrete diffusion lacks a counterpart to the continuous-domain probability-flow ODE, making it impossible to directly construct consistency models. The authors propose using the exact closed-form posterior bridge as a "stochastic PF-ODE surrogate" in the discrete domain to construct the Multi-Path Discrete Consistency (MPDC) training objective. This requires the denoiser's predictions across multiple stochastic bridge paths to be consistent in expectation. This enables the single-stage, teacher-free training of Consistent Diffusion Language Models (CDLM) capable of generating high-quality text in 2-3 steps, achieving SOTA in unconditional/conditional text generation and up to \(32\times\) speedup over AR models.
DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs: DAPD transforms the single-step parallel unmasking problem of dLLMs into a dynamic graph coloring problem of "selecting independent sets on self-attention-induced MRFs." Without any training, it simultaneously unmasks weakly dependent positions, reducing decoding steps to 1/3.87 of the original on LLaDA / Dream for multi-question mixed prompts with almost no loss in accuracy.
Degradation-Aware Metric Prompting for Hyperspectral Image Restoration: DAMP utilizes 6 interpretable spatial-spectral physical metrics (high-frequency energy ratio, texture uniformity, spectral curvature, etc.) as "Degradation Prompts" (DP) to replace black-box embeddings and explicit degradation labels. These DPs act as gating signals driving a Spatial-Spectral Adaptive MoE to select different "spatial/spectral experts," achieving SOTA performance across 5 HSI restoration tasks and 2 unseen degradations (motion blur, Poisson noise) simultaneously.
DyLLM: Efficient Diffusion LLM Inference via Saliency-based Token Selection and Partial Attention: DyLLM is a training-free inference acceleration framework for diffusion LLMs. It identifies "salient tokens" by measuring the cosine similarity of attention contexts between adjacent denoising steps. By recalculating FFN and attention only for these tokens using salient-aware approximate attention, it increases throughput to 7.6× / 9.6× on LLaDA / Dream with negligible performance loss.
Early Decisions Matter: Proximity Bias and Initial Trajectory Shaping in Non-Autoregressive Diffusion Language Models: This paper systematically characterizes the failure mechanism of masked diffusion language models (dLLM) under fully non-autoregressive (NAR) decoding. It identifies that proximity bias causes confidence-based sampling to degenerate into reverse autoregrssion, which is prematurely saturated by EOS tokens. By using a 5M-parameter lightweight planner and EOS temperature annealing to intervene in unmasking positions only at the first step, the authors improve LLaDA 8B NAR decoding by 2.8–4.3 points on reasoning tasks like GSM8K with almost no additional overhead.
From 2D Grids to 1D Tokens: Reforming Shared Representations for Multimodal Image Fusion: Multimodal image fusion has long relied on shared representations in 2D feature grids, leading to the entanglement of global appearance (brightness/contrast/tone) and local details, making them difficult to regulate independently. This paper moves "global appearance" into the compact token space of a frozen 1D tokenizer (TiTok-32). By employing "Selective Token Editing (STE)" to modify only a few token-channel entries, the method regulates global consistency while preserving a 2D pathway for detail recovery, achieving comprehensive SOTA results across four benchmarks.
Learning Normalized Energy Models for Linear Inverse Problems: The authors reformulate "linear inverse problems" as "anisotropic denoising" and propose Anisotropic Covariance Score Matching (A-CSM) to train a normalized energy model \(U_\theta(\mathbf{y},\boldsymbol{\Sigma})\approx -\log p(\mathbf{y}|\boldsymbol{\Sigma})\). A single model can handle inpainting, deblurring, and super-resolution while unlocking three new capabilities: energy-guided adaptive scheduling, MALA unbiased correction, and blind inverse estimation.
Measurement-Consistent Langevin Corrector for Stabilizing Latent Diffusion Inverse Problem Solvers: This paper reinterprets the instability of Latent Diffusion Model (LDM) inverse problem solvers as "solver dynamics deviating from the time-marginal distributions learned by the diffusion model." It proposes the MCLC module—a plug-and-play component that inserts a Langevin correction step, constrained to the orthogonal complement of the measurement gradient, after the measurement consistency step. This pulls the latent variables back to stable reverse diffusion trajectories without compromising measurement fidelity. MCLC consistently improves baselines such as LDPS, PSLD, and ReSample across various linear and non-linear degradation tasks on FFHQ and ImageNet.
One-shot Conditional Sampling: MMD meets Nearest Neighbors: CGMMD utilizes \(k\)-nearest neighbor graphs to estimate the "Expected Conditional MMD (ECMMD)" as a directly minimizable non-adversarial objective. It trains a conditional generator capable of sampling from \(P_{Y\mid X}\) in a single forward pass and provides non-asymptotic error bounds alongside proofs of distributional convergence.
One-Step Residual Shifting Diffusion for Image Super-Resolution via Distillation: To address the slow inference of diffusion SR models, this paper proposes RSD (Residual Shifting Distillation): distilling a 15-step ResShift teacher into a one-step student generator. The core mechanism involves "training the student so that a 'fake ResShift' trained on its output exactly matches the true teacher"—which is equivalent to matching the joint distribution (rather than just marginals as in VSD) of the teacher and student across all timesteps. Consequently, RSD outperforms the teacher and the comparable distillation method SinSR on LPIPS / CLIPIQA / MUSIQ. With only 174M parameters, 0.5GB VRAM, and 5 GPU-hours of training, it approaches the perceptual quality of massive T2I-based SR models.
Phy-CoSF: Physics-Guided Continuous Spectral Fields Reconstruction and Super-Resolution for Snapshot Compressive Imaging: Ours designs a train-render two-stage, wavelength-arbitrary queryable deep unfolding framework for Coded Aperture Snapshot Compressive Imaging (CASSI). By embedding a Continuous Spectral Field (CoSF) prior module—comprising a Fourier-Mamba-driven triple-branch cross-domain feature mixer, random frequency encoding, and a spectral synthesis head—within each unfolding stage, the model can be trained on discrete wavelengths and synthesize hyperspectral images at any continuous wavelength during inference, achieving continuous spectral reconstruction and zero-shot spectral super-resolution.
Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models: This paper proposes the Dilated Unmasking Scheduler (DUS): it uses predefined "equidistant gaps" to determine the unmasking order independent of model confidence. This reduces the number of denoiser calls per block of \(B\) tokens from \(\mathcal O(B)\) to \(\mathcal O(\log B)\), achieving a 5.8× wall-clock speedup on LLaDA / Dream / DiffuCoder while outperforming confidence-based parallel planners in quality.
PODiff: Latent Diffusion in Proper Orthogonal Decomposition Space for Scientific Super-Resolution: PODiff moves the diffusion process from pixel space to a fixed, variance-sorted POD coefficient space. By utilizing a minimal MLP, it achieves accuracy comparable to pixel-level diffusion on \(640\times 480\) SST downscaling tasks. Since reconstruction is linear, ensemble variance can be analytically back-propagated to physical space via \(\Sigma_u=\Phi\Sigma_a\Phi^\top\), yielding spatially interpretable and well-calibrated uncertainty.
Semi-Supervised Neural Super-Resolution for Mesh-Based Simulations: SuperMeshNet employs two complementary MPNNs—a primary model predicting LR→HR and an auxiliary model predicting HR-HR differences corresponding to LR-LR pairs—to mutually generate pseudo-labels for unpaired samples. Combined with two lightweight inductive biases (node-level and message-level centering), this approach allows PDE mesh super-resolution to outperform a 100% HR fully supervised baseline using only 10% HR data, consistently reducing RMSE across six MPNN architectures.
Solving Inverse Problems with Flow-based Models via Model Predictive Control: This work reformulates "solving inverse problems with pre-trained flow models" as Model Predictive Control (MPC)—instead of optimizing the entire sampling trajectory at once, it solves a short-horizon subproblem at each time step, applies one step of control, and re-plans. This significantly reduces memory usage and leads to a variant that requires no backpropagation through the flow model, enabling the scaling of training-free guidance to quantized FLUX.2 (32B) models on consumer-grade 24GB GPUs.
Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges: SDB reframes modality translation as "selecting a coupling from the set \(\mathcal{P}\) of all couplings satisfying marginal constraints." Built upon LDDBM, it incorporates marginal matching (WTA + capacity constraints) and dual-layer cycle consistency (endpoint and trajectory levels). Paired supervision is treated as an optional heuristic, enabling the model to function under zero, semi, and fully paired budgets. Even under full supervision, it outperforms paired-only baselines (e.g., FFHQ→CelebA-HQ PSNR increases from 25.6 to 25.9).
Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems: Optimizing Guidance and Stochasticity Schedules: This paper systematically treats the three forces long considered constants in diffusion posterior sampling—Data Consistency (DC) guidance, Classifier-Free Guidance (CFG), and stochasticity—as a coupled time-varying triadic system for the first time. It provides theoretical and empirical proof that early-stage CFG conflicts with DC directions, while stochasticity pulls trajectories back toward high-probability manifolds. Based on this, it proposes a monotonic triadic scheduling trend of "DC↓, CFG↑, η↓" and utilizes "Template Search + GRPO Reinforcement Learning" to find optimal curves, simultaneously refreshing distortion and perceptual metrics in super-resolution and deblurring on FFHQ and DIV2K.
UOTIP: Unbalanced Optimal Transport Mapping for Unpaired Inversion Problems: The UOTIP method is proposed—formulating the unpaired image inversion problem as mapping learning from a noisy measurement distribution to a clean signal distribution through an Unbalanced Optimal Transport (UOT) framework, achieving robustness and theoretical guarantees by introducing a likelihood cost function and a quadratic cost term.