ECCV 2026 论文列表

Modern convolutional neural networks (CNNs) organize com-putation as a discrete stack of layers whose parameters are independentlystored and learned, with the number of layers fixed as an architecturalhyperparameter. In this work, we explore an alternative perspective: cannetwork parameterization itself be modeled as a continuous dynamicalsystem? We introduce Puppet-CNN, a framework that represents con-volutional layer parameters as states evolving along a learned parameterflow governed by a neural ordinary differential equation (ODE). Underthis formulation, layer parameters are generated through continuousevolution in parameter space, and the effective number of generatedlayers is determined by the discretization resolution used to sample thelearned parameter trajectory, which can be modulated by input com-plexity to enable input-adaptive computation. We validate its feasibilityon standard image classification benchmarks and show that continuousparameter dynamics can maintain competitive predictive performancewhile substantially reducing stored trainable parameters. These resultssuggest that viewing neural network parameterization through the lensof dynamical systems provides a structured and flexible design space foradaptive convolutional models.

Self-supervised learning methods learn high-quality visualrepresentations, yet recent studies show that these representations of-ten capture demographic biases present in the training data. Existingfairness-aware methods address this by redesigning the self-supervisedobjective itself, limiting portability across the rapidly evolving land-scape of self-supervised learning (SSL) frameworks. We propose Proto-Fair, a fairness-aware contrastive loss designed to work alongside ex-isting SSL objectives without modifying them. ProtoFair leverages un-supervised prototype clustering to identify pseudo-counterfactual pairs:samples sharing the same cluster assignment but belonging to differentsensitive groups. By pulling these content-matched, cross-group samplestogether in the embedding space, ProtoFair encourages the encoder tolearn representations that are invariant to the sensitive attribute. Themethod requires only sensitive attribute annotations, no target labels,and integrates seamlessly with both SimCLR and SupCon. Experimentson CelebA and UTKFace demonstrate consistent fairness improvementswhile maintaining competitive accuracy.

Image Quality Assessment (IQA) requires aligning modeloutputs with human perception. While Vision-Language Models (VLMs)show promise, current methods struggle with generalization due to alack of fine-grained reasoning alignment. Existing RL-based approachesprimarily rely on outcome rewards; however, their inherent sparsity oftenresults in accurate predictions based on erroneous logical attributions.This phenomenon creates a critical bottleneck that stifles the develop-ment of robust reasoning capabilities and limits performance gains inout-of-distribution scenarios. To address this problem, we propose RGPR,which utilizes grounding process rewards to orchestrate VLM trajectoriesalong a human-emulated reasoning chain. To maintain optimization sta-bility, we introduce step-wise reward normalization to preserve trainingsignal precision. RGPR achieves state-of-the-art results across three tasks,demonstrating superior reasoning fidelity and robust generalization.

Forecasting the future anatomy of slow-evolving neurode-generative diseases could enable earlier, more targeted intervention andimprove clinical trial design, but it remains challenging because true pro-gression signals are subtle in longitudinal MRI. In this low-signal regime,transferring modern generative sequence models directly is unreliable:training is dominated by stable baseline anatomy and confounded bydense, sample-specific nuisance variation. We first provide a theoreticalanalysis that explains these failures through two modes. Identity col-lapse occurs when optimization is driven toward reproducing the currentanatomy, which prevents the model from learning faint temporal change.The continuous interpolation trap arises when standard smooth networkscannot separate localized biological drift from pervasive noise, whichleads to spurious changes that diffuse across the volume. To addressboth issues, we propose Latent Drift, a progressive generative frame-work that learns change in a compressed semantic representation ratherthan synthesizing full-resolution anatomy. This design removes pixel-levelidentity from the prediction target and concentrates model capacity onprogression-relevant dynamics. We further apply Finite Scalar Quan-tization to the learned change representation, which suppresses small,high-frequency nuisance fluctuations while preserving consistent struc-tural drift. Experiments on longitudinal 3D brain MRI show that LatentDrift improves patient-specific neuro-forecasting over diffusion and au-toregressive transformer baselines across generative fidelity and clinicallyrelevant evaluation metrics.

Capturing high-speed polarization dynamics is crucial forrevealing physical cues like stress variations and surface structures thatare invisible in conventional intensity imaging. While snapshot compres-sive imaging (SCI) enables high-speed video reconstruction from a singlemeasurement, extending it to division-of-focal-plane (DoFP) polarizationsensing introduces a fundamentally more ill-posed inverse problem due tothe coupling between temporal multiplexing and spatial polarization mo-saicing. To address this challenge, we propose Polarization Coded Aper-ture Compressive Temporal Imaging (Pol-CACTI), a hardware–softwareco-designed framework for polarized video SCI. First, we build an opticalsystem that encodes DoFP polarization dynamics into a single snapshotthrough spatiotemporal modulation and integration. Second, we con-struct a large-scale high-quality polarized video dataset with carefullydenoised Stokes-domain supervision, enabling supervised learning of po-larization video SCI reconstruction models. Third, we develop a physics-driven reconstruction network that jointly performs temporal recoveryand polarization demosaicing, where the Polarization-Aware LearnableBack-Projection (PA-LBP) module is proposed to enforce measurementconsistency with the raw DoFP mosaic and mitigates error accumula-tion caused by decoupled reconstruction pipelines. Extensive experimentsdemonstrate that our method significantly improves both intensity fi-delity and polarization estimation accuracy. The dataset and code arereleased to https://github.com/kaeoqxhailesxya/Pol-CACTI.

Reconstructing 4D scenes from casually captured monocularvideo is vital for applications in virtual reality (VR) and embodied AI.Recent advances in 4D reconstruction and novel view synthesis havesubstantially propelled this capability. However, existing reconstructionmethods generally cannot recover regions beyond visible camera limits.Consequently, we introduce a new paradigm that achieves 4D scene syn-thesis by combining visible-region reconstruction from monocular inputwith invisible-region generation beyond observable camera boundaries.A straightforward approach is to leverage video generation models as“generative priors” for invisible-region exploration. However, their inherentstochasticity and large solution space prevent stable and consistent viewsynthesis. As a result, naively incorporating generative content into thereconstruction process often causes artifacts, especially when camera tra-jectories deviate significantly from the original video. To overcome thesechallenges, we present Panoramic Alignment for Strategic Exploitation ofGenerative Priors (PASTEL). Specifically, PASTEL proposes panoramicscene alignment, a novel representation that reformulates the intractable3D "invisible region” exploration into a tractable 2D directional trajectoryplanning. This is achieved by reducing the viewpoint planning from 6-DoFsearch to a 2D directional search with explicit visibility boundaries. Byoperating within this panoramic space, our method strategically iden-tifies camera trajectories that maximize exploration beyond observableboundaries while minimizing viewpoint deviation. Experimental resultsshow that PASTEL can not only extrapolate plausible scene contentbeyond the observable boundaries of input monocular videos, but alsosubstantially boost monocular 4D reconstruction performance. PASTELoutperforms the previous state-of-the-art method by 0.9dB in full-imagePSNR on the DyCheck IPhone dataset.

Reasoning video object segmentation predicts pixel-level masksin videos from natural-language queries that may involve implicit andtemporally grounded references. However, existing methods are devel-oped and evaluated in an offline regime, where the entire video is availableat inference time and future frames can be exploited for retrospective dis-ambiguation, deviating from real-world deployments that require strictlycausal, frame-by-frame decisions. We study Online Reasoning Video Ob-ject Segmentation (ORVOS), where models must incrementally interpretqueries using only past and current frames without revisiting previouspredictions, while handling referent shifts as events unfold. To supportevaluation, we introduce ORVOSB, a benchmark with frame-level causalannotations and referent-shift labels, comprising 210 videos, 12,907 anno-tated frames, and 512 queries across five reasoning categories. We furtherpropose a baseline with continually-updated segmentation prompts anda structured temporal token reservoir for long-horizon reasoning underbounded computation. Experiments show that existing methods struggleunder strict causality and referent shifts, while our baseline establishesa strong foundation for future research.

Reinforcement learning for diffusion models primarily relieson on-policy learning to optimize the model by using samples from itsown current policy. Despite offering a direct optimization signal, theapproach is fundamentally self-limiting and prone to reward overfittingand mode collapse due to its myopic guidance. In this paper, we pro-pose a new framework, namely DiffusionCompass, that breaks such lim-itation by strategically integrating off-policy guidance. DiffusionCom-pass establishes a dual learning objective: the model continues to ex-plore reward maximization on-policy, while simultaneously being an-chored by off-policy examples that define a stable quality benchmark.That is achieved through a novel integration of several key techniques:an off-policy objective with distribution-aware filtering to adaptively em-phasize relevant anchors, a perceptual anchored reward mechanism tomitigate reward hacking, and a reward renormalization strategy usingoff-policy samples for a stable quality baseline. The dual regimes com-plement each other: off-policy guidance prevents reward overfitting andcollapse, and on-policy exploration ensures the model does not merelyimitate a static dataset. Experiments across text-to-image/video gener-ation demonstrate that DiffusionCompass achieves superior reward per-formance while maintaining high sample diversity, shaping a robust andeffective path for fine-tuning generative models. Code is available athttps://github.com/HiDream-ai/DiffusionCompass.

Recent joint audio-visual diffusion models achieve remark-able generation quality but suffer from high latency due to their bidirec-tional attention dependencies, hindering real-time applications. We pro-pose OmniForcing, the first framework to distill an offline, dual-streambidirectional diffusion model into a high-fidelity streaming autoregressivegenerator. However, naively applying causal distillation to such dual-stream architectures triggers severe training instability, due to the ex-treme temporal asymmetry between modalities and the resulting tokensparsity. We address the inherent information density gap by introducingan Asymmetric Block-Causal Alignment with a zero-truncation GlobalPrefix that prevents multi-modal synchronization drift. The gradient ex-plosion caused by extreme audio token sparsity during the causal shift isfurther resolved through an Audio Sink Token mechanism equipped withan Identity RoPE constraint. Finally, a Joint Self-Forcing Distillationparadigm enables the model to dynamically self-correct cumulative cross-modal errors from exposure bias during long rollouts. Empowered by amodality-independent rolling KV-cache inference scheme, OmniForcingachieves state-of-the-art streaming generation at ∼25 FPS on a singleGPU, maintaining multi-modal synchronization and visual quality on parwith the bidirectional teacher. Project Page: https://omniforcing.com.

Recent works in novel view synthesis, e.g., Neural RadianceField (NeRF) and 3D Gaussian Splatting (3DGS), have significantly ad-vanced rendering quality and efficiency. However, existing Gaussian-basednovel view synthesis methods typically follow a single-view optimizationparadigm. We observed that this optimization paradigm suffers fromunstable gradients, leading to suboptimal rendering quality. To tackle thisissue, we present a novel multi-view regulated Gaussian Splatting (MVGS)that fully leverages a multi-view coherent (MVC) constraint throughoutthe optimization process. Specifically, our proposed MVC enhances 3DGaussian multi-view consistency and thus ensures smoother gradient up-dates. Furthermore, since single-scale training usually leads to suboptimalsolutions, we propose a cross-intrinsic guidance scheme in a coarse-to-fine manner to improve the convergence of multi-view optimization in3DGS. In particular, by incorporating more multi-view images at the lowresolution, we can optimize 3D Gaussians with more comprehensive per-spectives. Then, finer-scale Gaussians are initialized by coarsely estimatedones instead of optimizing full-scale 3D Gaussians from scratch. Moreover,we found that 3D Gaussians usually struggle to fit 2D training views withminimal overlap. Thus, we propose a novel multi-view cross-ray densifica-tion strategy, where 3D Gaussians are dynamically split to accommodatedrastic viewpoint variations in the multi-view optimization process. Inthis way, the multi-view consistency can be further improved. Notably,our proposed MVGS method is a plug-and-play optimizer. Extensiveexperiments across various tasks demonstrate that our proposed MVGSimproves existing Gaussian-based methods and achieves state-of-the-artperformance. Project Page: https://xiaobiaodu.github.io/mvgs-project/

Embodied agents operating in the real world require multi-scale reasoning and knowledge adaptation as conditions change. We iden-tify two challenges in applying Mixture of Experts (MoE) to this setting:routing lacks an explicit notion of scale, preventing targeted updates atspecific scales, and a uniform update policy cannot accommodate thedifferent rates at which knowledge at each scale becomes outdated. Wepresent MuSix, a framework that addresses both challenges throughscale-aware world model mixture and evolution. A two-stage routingmechanism first maps experiential distance, a measure of situational nov-elty inspired by Construal Level Theory, to a weight over continuous scalespace via a meta-router, then selects world models within the identifiedscale. For adaptation, scale-dependent forgetting rates allow low-scaleknowledge to refresh rapidly while high-scale abstractions persist, andgated inter-scale transfer maintains coherence across the hierarchy. Ex-periments on EmbodiedBench and HAZARD show that MuSix improvesover state-of-the-art baselines.

We present MGM-Omni, an Omni MLLM for omni-modalunderstanding and expressive, long-horizon speech generation. MGM-Omni adopts a “brain–mouth” design with a dual-track, token-based ar-chitecture that cleanly decouples multimodal reasoning from real-timespeech generation. This design enables efficient cross-modal interactionand low-latency, streaming speech generation. For multimodal under-standing, a unified training strategy coupled with a dual audio encoderdesign enables long-form omni modal perception across diverse acous-tic conditions. For speech generation, a chunk-based parallel decodingscheme narrows the text–speech token-rate gap, accelerating inferenceand supporting streaming zero-shot voice cloning with stable timbre overextended durations. Compared to concurrent work, MGM-Omni achievesthese capabilities with markedly data-efficient training. Extensive exper-iments demonstrate that MGM-Omni outperforms existing open sourcemodels in preserving timbre identity across extended sequences, produc-ing natural and context-aware speech, and achieving superior long-formaudio and omnimodal understanding. MGM-Omni establishes an effi-cient, end-to-end paradigm for omnimodal understanding and control-lable, personalized long-horizon speech generation. Code and models areavailable at https://github.com/dvlab-research/MGM-Omni.

Existing data pruning methods predominantly rely on high-dimensional feature embeddings to measure sample importance. How-ever, these compressed vectors often obscure fine-grained semantic inter-actions, leading to suboptimal coverage of rare semantic concepts in thepruned subsets. In this paper, we propose Mapping the Concept Land-scape (MCL), a novel structural perception framework for transparentdata pruning. Instead of abstract embeddings, we represent each image-caption pair as an explicit sample-level graph comprising entities, events,and attributes. By integrating these individual graphs into a comprehen-sive dataset-level graph, we characterize the global distribution of seman-tic concepts and quantify their rarity across the entire corpus. Basedon this structured perception, we develop a greedy concept-coveragemaximization algorithm that iteratively selects samples to maximize themarginal gain of high-value, under-represented concepts. Experimentalresults on various benchmarks demonstrate that our method not onlyachieves superior pruning efficiency compared to state-of-the-art meth-ods but also provides a transparent and interpretable audit trail for theselection process.

Generative world models for autonomous driving (AD) are ofgreat value in applications such as data augmentation, closed-loop sim-ulation, and safety-critical scenario evaluation. Unlike the widely stud-ied image modality, in this work we explore generative world modelsfor LiDAR data. Existing generation methods for LiDAR predominantlyfocus on single frame generation or lack the capacity for interactive sim-ulation, while existing prediction approaches require multiple frames ofhistorical input and can only deterministically predict multiple framesat once. Both paradigms fail to support long-horizon interactive gen-eration. To this end, we introduce LaGen, which, to the best of ourknowledge is the first autoregressive framework capable of generatinglong-horizon LiDAR scenes in a frame-by-frame, interactive manner. La-Gen is able to take a single-frame input as a starting point and effectivelyutilize bounding box information as conditions to generate high-fidelity4D scene. In addition, we introduce a scene decoupling estimation moduleto enhance the model’s interactive generation capability for object-levelcontent, as well as a noise modulation module to mitigate error accumu-lation during long-horizon generation. We extensively evaluate LaGen’sperformance in controlled data generation and long-horizon scene genera-tion on the nuScenes dataset. The experimental results demonstrate thatLaGen achieves state-of-the-art performance, especially on later frames.The code is publicly available at: https://github.com/szzhou88/LaGen.

Recent advancements in 3D generation have significantly en-hanced the fidelity and geometric details of synthesized 3D assets. How-ever, due to the inherent ambiguity of single-view input and the lackof robust global structural priors–caused by limited 3D training data–the unseen regions generated by existing models remain stochastic anddifficult to control. This often results in geometries that are either phys-ically implausible or misaligned with user intent. In this paper, we pro-pose Know3D, a novel framework designed to incorporate rich knowl-edge from Multimodal Large Language Models (MLLMs) into 3D gen-eration processes. By leveraging latent hidden-state injection, Know3Dsupports language-controllable generation of the back-view for 3D as-sets. We utilize a VLM-diffusion-based architecture: the Vision LanguageModel (VLM) is used to provide high-level semantic understanding, whilethe diffusion model serves as a bridge, transferring semantic knowledgeinto the 3D generation model. Extensive experiments demonstrate thatKnow3D effectively bridges the gap between abstract textual instruc-tions and the geometric reconstruction of invisible regions. By trans-forming the traditionally stochastic back-view hallucination into a se-mantically controllable process, Know3D offers a promising direction forhighly plausible and user-friendly 3D generation in the future. Projectpage: https://xishuxishu.github.io/Know3D.github.io/.

Text-to-image (T2I) diffusion models often fail to faithfullyrender explicit textual descriptions, instead defaulting to strongly learnedvisual priors due to a phenomenon referred to as concept association bias.We show that such bias is particularly strong for one-and-only (OAO)objects, entities that exist in a single canonical form, such as celestialbodies, landmarks, and artworks. The deeply ingrained visual identity forthese concepts often resists modification through prompting alone. Ad-dressing this challenge, we first identify through an information-theoreticanalysis that the final text embedding discards concept-level informationpresent in the intermediate-layer text representations, reducing the mu-tual information available to the subsequent denoising process. We thenpropose Intermediate Text Representation (IR)-guided diffusion, whichinjects intermediate hidden states of the text encoder into the condi-tioning signal during early denoising steps, recovering suppressed con-cepts without any additional training, optimization, or external models.To systematically evaluate the challenging task of aligning generativeoutputs with unusual prompts for OAO objects, we introduce OAO-AttackBench, a benchmark comprising counterfactual prompts that di-rectly conflict with the core visual identity of OAO objects. Experi-ments on four benchmarks, including OAO-AttackBench, show that ourmethod achieves up to a 19.1 percentage-point improvement in VQAS-core while preserving generation fidelity and human preference. Projectpage: https://soyoun-won.github.io/one-and-only-ir-guidance/.

Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models despite their strong text-to-image generation capabilities at limited resolutions. In this work, we find two key challenges particularly hindering the application of offthe-shelf DiT models for high-resolution image synthesis in a training-free manner, namely, spatial disorder and long generation time. To address these challenges, we propose a novel method tailored to adapt off-theshelf DiT models for high-resolution image synthesis. Extensive experiments show the efficacy of our method.

Open-vocabulary 3D object detection aims to localize andrecognize objects beyond a fixed training taxonomy. In multi-view RGBsettings, recent approaches often decouple geometry-based instance con-struction from semantic labeling, generating class-agnostic fragments andassigning open-vocabulary categories post hoc. While flexible, such de-coupling leaves instance construction governed primarily by geometricconsistency, without semantic constraints during merging. When geomet-ric evidence is view-dependent and incomplete, this geometry-only merg-ing can lead to irreversible association errors, including over-mergingof distinct objects or fragmentation of a single instance. We proposeGroup3D, a multi-view open-vocabulary 3D detection framework thatintegrates semantic constraints directly into the instance constructionprocess. Group3D maintains a scene-adaptive vocabulary derived from amultimodal large language model (MLLM) and organizes it into semanticcompatibility groups that encode plausible cross-view category equiva-lence. These groups act as merge-time constraints: 3D fragments are asso-ciated only when they satisfy both semantic compatibility and geometricconsistency. This semantically gated merging mitigates geometry-drivenover-merging while absorbing multi-view category variability. Group3Dsupports both pose-known and pose-free settings, relying only on RGBobservations. Experiments on ScanNet and ARKitScenes demonstratethat Group3D achieves state-of-the-art performance in multi-view open-vocabulary 3D detection, while exhibiting strong generalization in zero-shot scenarios. The project page is available at https://ubin108.github.io/Group3D/.

Pre-trained Latent Diffusion Models (LDMs) have recentlyshown strong perceptual priors for low-level vision tasks, making thema promising direction for multi-exposure High Dynamic Range (HDR)reconstruction. However, directly applying LDMs to HDR remains chal-lenging due to: (1) limited dynamic-range representation caused by 8-bitlatent compression, (2) high inference cost from multi-step denoising,and (3) content hallucination inherent to generative nature. To addressthese challenges, we introduce GMODiff, a gain map-driven one-stepdiff usion framework for multi-exposure HDR reconstruction. Instead ofreconstructing full HDR content, we reformulate HDR reconstructionas a degradation-aware Gain Map (GM) refinement problem, where theGM encodes the extended dynamic range while retaining the same bitdepth as LDR images. We initialize the denoising process from an in-formative regression-based estimate rather than pure noise, allowing themodel to generate high-quality GMs in a single denoising step. Further-more, recognizing that regression-based models excel in content fidelitywhile LDMs favor perceptual quality, we leverage regression priors toguide both the denoising process and latent decoding of the LDM, sup-pressing hallucinations while preserving structural accuracy. Extensiveexperiments demonstrate that our GMODiff performs favorably againstseveral state-of-the-art methods and is 100× faster than previous LDM-based methods.