ECCV 2026 Papers

On-the-fly category discovery requires deciding for each in-coming test sample whether to assign it to an existing category or spawna new one. Existing methods typically implement this decision throughmatching-based heuristics, such as radius- or hash-based rules. While ef-fective in practice, these methods usually treat category birth implicitlyas a fallback when no existing category matches confidently, rather thanas an explicit alternative supported by its own statistical evidence. Toaddress this, we propose DP-BOA, a posterior-predictive decision frame-work based on an online Dirichlet-process Gaussian mixture model witha Normal–Inverse–Wishart prior. During training, we use labeled datato calibrate a shared NIW prior over category Gaussians and warm-startthe known-category posteriors. At test time, for each incoming sample,DP-BOA compares the posterior predictive evidence for assignment toexisting categories against the evidence for spawning a new category in-duced by the DP prior, and then updates category statistics online afterthe decision. The method captures anisotropic category geometry andnaturally adapts decision confidence as evidence accumulates. Acrossstandard OCD benchmarks, DP-BOA consistently outperforms strongbaselines and delivers particularly strong novel-class discovery perfor-mance while maintaining competitive known-class accuracy. The projectpage is available at DP-BOA.

Pose-agnostic Anomaly Detection (PAD) remains challeng-ing as anomalies can appear under arbitrary viewpoints, requiring meth-ods to handle significant pose variations. Existing approaches rely oncomplex 3D reconstruction, which are computationally expensive and re-quire extensive multi-view data. We propose PADFormer, a novel image-space approach that leverages Vision Transformer (ViT) to directly re-construct anomaly-free versions of query images while preserving poseinformation. Our key insight is to adapt cross-view masked reconstruc-tion for anomaly detection through training exclusively on normal data,combined with dynamic patch selection and spatial alignment mecha-nisms that enable effective learning from sparse reference views undersignificant pose variations. During inference, we perform multiple for-ward passes with different masking patterns to generate an ensembleof anomaly-free reconstructions, ensuring comprehensive coverage of thequery image. Anomalies are detected by comparing these reconstructionswith the query image. PADFormer achieves state-of-the-art results on thePAD benchmark while maintaining comparable performance on classicfew-shot anomaly detection (FSAD) tasks, demonstrating superior effi-ciency and generalization without requiring 3D reconstruction.

Rapid advances in diffusion models have raised concernsabout privacy, harmful content, and copyright violations. Concept era-sure methods aim to address these risks by removing undesirable con-cepts from pretrained models without full retraining. However, existingapproaches often degrade non-target generation quality through heuristicstrength scaling (e.g., classifier-free guidance or negative prompt scaling),leading to uncontrolled and excessive semantic drift. To address this, wepropose Push–Pull Attentional Anchoring (PPAA), a mechanism in thecross-attention space that displaces target representations while anchor-ing the scale of erasure via bounded cosine-similarity constraints. Byusing a relative similarity ratio instead of an absolute difference, ouradaptive threshold represents a meaningful percentage of semantic re-tention, facilitating a single value to be applied uniformly across layers.We conduct extensive experiments across multiple concept categories, in-cluding celebrities, artistic styles, nudity, and intellectual property. Ourmethod achieves favorable trade-offs against prior approaches on StableDiffusion v1.4 and demonstrates its applicability to modern architecturessuch as FLUX.

Diffusion-based video super-resolution (VSR) methods deliver strong perceptual quality but are often unsuitable for latency-sensitive scenarios due to reliance on future frames and expensive multi-step denoising. We propose Stream-DiffVSR, a causally conditioned diffusion framework for efficient online VSR. Operating strictly on past frames, Stream-DiffVSR integrates a four-step distilled denoiser for fast inference, an Auto-regressive Temporal Guidance (ARTG) module that injects motion-aligned cues during latent denoising, and a lightweight temporalaware decoder with a Temporal Processor Module (TPM) to enhance detail and temporal coherence. Unlike chunk-wise streaming inference, our strictly frame-by-frame causal design avoids sequence-level waiting, substantially reducing time-to-first-frame and end-to-end latency. Stream-DiffVSR processes 720p frames in 0.328 seconds on an RTX 4090 and consistently outperforms prior diffusion-based baselines. Compared with the online state-of-the-art TMP, it improves perceptual quality (LPIPS +0.095). Compared with prior diffusion-based VSR methods such as MGLD-VSR [104], it reduces per-frame runtime by over 130×. Moreover, Stream-DiffVSR substantially lowers time-to-firstframe for diffusion-based VSR, reducing initial delay from over 4600 seconds to 0.328 seconds, making diffusion-based VSR markedly more practical for low-latency online and streaming deployment. Project page: https://jamichss.github.io/stream-diffvsr-project-page/

We propose a remeshing-enhanced method for generativelydeforming shapes with visual losses. It is intuitive that sufficiently dras-tic deformations of a mesh without changing its triangulation can easilycompromise element quality, even if such large geometry changes maybe semantically desired. Shape deformation methods could thus bene-fit from changing the triangulation; however, this is not done by mostgenerative, text-based, visually-supervised mesh deformation methods.Remeshing is a discrete operation, proven to be especially challengingto couple with the notoriously noisy supervision signal provided by vi-sual losses. We propose a vertex-based deformation optimization quantitycapable of large deformations and robustness to such noise; we period-ically remesh using an isotropic remesher that interpolates and carriesforward the deformation optimization state. This enables continuous,geometry-informed progress in coarse-to-fine addition of resolution. Theresulting shapes’ triangulations fit their optimized geometry and haveneat isotropic elements. Further, our method is localizable, able to grownew features on a base shape with expressive detail, leaving the rest un-changed. We showcase the effectiveness of our method on a variety ofshapes and prompts, both local and global deformations, and demon-strate its superior visual quality and triangle efficiency. Our project pageis at https://threedle.github.io/radmesh.

Neural implicit representations have recently achieved im-pressive results in novel view synthesis and multi-view 3D reconstruc-tion, yet both NeRF- and Gaussian Splatting-based methods require per-scene optimization, which makes them inefficient. Generalizable NeuralSurface Reconstruction (GNSR) methods have been proposed to removethis need by learning feature representations directly predicted from in-put images. However, their typical reliance on dense feature volumesseverely limits achievable resolution and fidelity due to prohibitive mem-ory costs. We introduce Sparse Volumetric Reconstruction (SVRecon),a new GNSR framework that unlocks high-resolution, memory-efficientreconstruction through learned occupancy-driven sparsity, in a more ef-fective way than earlier approaches to introducing sparsity in GNSRs.Our approach uses a nested two-stage architecture: (1) an occupancyprediction network that identifies surface-containing voxels, and (2) ahigh-resolution sparse volume rendering framework defined only withinthese occupied regions, together with specialized sparsified algorithmsfor ray sampling, feature aggregation, and querying. This design enablesfine-grained surface reconstruction while avoiding the heavy memoryfootprint of dense grids. SVRecon operates at resolutions up to 5123on standard 32GB hardware—substantially higher than prior generaliz-able methods—and delivers smoother and more precise reconstructionsacross diverse datasets, particularly in sparse-view settings.

Recent progress in video diffusion models has spurred grow-ing interest in camera-controlled novel-view video generation for dy-namic scenes, aiming to provide creators with cinematic camera con-trol capabilities in post-production. A key challenge in camera-controlledvideo generation is ensuring fidelity to the specified camera pose, whilemaintaining view consistency and reasoning about occluded geometryfrom limited observations. To address this, existing methods either traintrajectory-conditioned video generation model on trajectory–video pairdataset, or estimate depth from the input video to reproject it alonga target trajectory and generate the unprojected regions. Nevertheless,existing methods struggle to generate camera-pose–faithful, high-qualityvideos for two main reasons: (1) reprojection-based approaches are highlysusceptible to errors caused by inaccurate depth estimation; and (2)the limited diversity of camera trajectories in existing datasets restrictslearned models. To address these limitations, we present InfCam, a depth-free, camera-controlled video-to-video generation framework with highpose fidelity. The framework integrates two key components: (1) infi-nite homography warping, which encodes 3D camera rotations directlywithin the 2D latent space of a video diffusion model. Conditioningon this noise-free rotational information, the residual parallax term ispredicted through end-to-end training to achieve high camera-pose fi-delity; and (2) a data augmentation pipeline that transforms existing syn-thetic multiview datasets into sequences with diverse trajectories and fo-cal lengths. Experimental results demonstrate that InfCam outperformsbaseline methods in camera-pose accuracy and visual fidelity, generaliz-ing well from synthetic to real-world data.

Square jigsaw puzzles are typically solved by visually match-ing piece images to recover the original layout. This work introducesPuzLM, an alternative perspective that recasts jigsaw reassembly asa discrete sequence-to-sequence (Seq2Seq) problem, inspired by naturallanguage representations. We design an efficient puzzle quantization pro-cedure that transforms each piece into a short sequence of discrete tokens,enabling the direct application of standard Seq2Seq language models aspowerful jigsaw solvers. Our approach demonstrates that accurate puzzlereconstruction can be achieved through purely symbolic reasoning overdiscrete representations, improving state-of-the-art performance even onpuzzles with eroded boundaries or missing pieces.

Path-based attribution methods such as Integrated Gradi-ents (IG) are widely adopted for their strong axiomatic properties andeffectiveness in attributing model predictions to input features by inte-grating gradients along a path from a baseline to the input. However,the choice of the attribution path largely affects the quality of explana-tions, and existing approaches rely on fixed or hand-crafted paths thatoften produce noisy or distorted attributions. To address this limitation,we propose Diffusion Integrated Gradients (DiffIG), a novel method thatreformulates path generation as a conditional generative modeling prob-lem. DiffIG first trains a diffusion model to learn a distribution overpaths generated from a Stick-Breaking Process, then employs guidedsampling to embed user guidance during the sampling procedure. Wedemonstrate that DiffIG quantitatively matches or outperforms existingpath-based methods, achieving perceptually aligned explanations. Thiswork introduces a new generative perspective for flexible, inference-timecontrollable Explainable Artificial Intelligence (XAI) methods.

How can we bridge the numerical chasm between full-precisionCNN models and highly quantized targets when original training dataare unavailable? As the demand for deploying deep neural networkson resource-constrained edge hardware grows, achieving extreme modelcompression without compromising the privacy of training data has be-come a critical challenge. However, direct transitions to ultra-low-bitregimes introduce extreme optimization instability and activation col-lapse, where high quantization noise causes deep layers to lose their dis-criminative capacity and representational power.In this paper, we propose Zest (ZSQ Enhancement via Stepping-stoneTransition), an accurate zero-shot quantization framework for CNNs thatbridges the precision gap between full-precision models and ultra-low-bit targets through a unified two-stage distillation and reconstructionpipeline. Rather than attempting a direct and unstable jump from fullprecision to an ultra-low bit quantized model, Zest introduces a stableintermediate bit stage to tame the numerical distribution and regular-ize the model. Furthermore, we propose a two-phase distillation strategythat initializes high-fidelity semantic skeletons from the full-precisionmodel and subsequently recalibrates them through an intermediate bitstage to ensure calibration data are numerically aligned with quantized-domain constraints. Experimental results demonstrate that this progres-sive transition leads to better distillation, achieving 70.50% top-1 accu-racy on ResNet-18 at W 4A4 precision and showing higher performancecompared to existing SOTA methods in various models.

Chinese font production is a complex task that balances vi-sual consistency, vector editability, and topological integrity at a massivecharacter scale. Existing fragmented toolchains, relying on manual rules,often lead to geometric redundancy and defects. This paper introducesFontCopilot, a multimodal large language model that performs raster-to-vector conversion, Bézier curve sparsi_x001C_cation, defect marking, and geo-metric correction within a single model. We elevate font engineering fromexplicit geometric rule constraints to structured glyph code understand-ing and rewriting, serializing Bézier curves into SVG XML. This enablesthe model to perform topological reorganization in the code space andinternalize implicit designer priors as structured reasoning constraints.FontCopilot integrates few-shot reference guidance and zero-shot gen-eralization mechanisms, balancing speci_x001C_c style transfer and universalstandard reconstruction. To advance the development of a uni_x001C_ed fontengineering model, we constructed the CFCopilot-8M dataset, covering300 font types and 8 million glyph samples. Under progressive multi-taskhybrid training, the glyphs processed by FontCopilot show consistentimprovements in editability, contour compactness, and defect correction,validating its industrial-grade font production value.

Facial privacy has become a critical concern as unauthorized large-scale image scraping enables malicious actors to collect personal face images without consent and exploit them for training unauthorized recognition systems. Protecting facial privacy requires modifying a face image such that automated recognition systems fail to identify the individual, while the visual appearance remains unchanged to human observers, ensuring the image retains its social utility. However, existing methods fail to satisfy both requirements; they either irreversibly degrade visual quality, rendering images impractical, or alter facial identity to such an extent that the protected image bears no resemblance to the original subject. To address this fundamental privacy-utility tradeoff, we propose MIRAGE (Multi-scale Identity Removal via Attention-Guided Encoding), a frequency-selective facial privacy framework grounded in the observation that identity-discriminative features reside predominantly in low-frequency image components, while visual appearance is encoded in high-frequency components. We further introduce FAC (FrequencyAware Consistency) Loss, which jointly enforces identity separation and visual fidelity during training. A comprehensive evaluation across three benchmark datasets and nine state-of-the-art (SOTA) deep face recognition (DFR) models demonstrates that MIRAGE achieves robust facial privacy protection while preserving visual appearance, successfully bridging the privacy-utility gap left unresolved by existing methods. The code is available at: https://github.com/atulkr05/MIRAGE.git.

Large Vision-Language Models (LVLMs) encounter significant challenges in image understanding and visual reasoning, leading to critical perception failures. Visual prompts, which incorporate image manipulation code, have shown promising potential in mitigating these issues. While visual prompts have emerged as a promising direction, previous methods for visual prompt generation have focused on tool selection rather than diagnosing and mitigating the root causes of LVLM perception failures. Because of the opacity and unpredictability of LVLMs, optimal visual prompts must be discovered through empirical experiments, which have relied on manual human trial-and-error. In this work, we propose an automated semantic exploration framework for discovering task-wise visual prompts. Unlike previous methods, our approach enables diverse yet efficient exploration through agent-driven experiments, minimizing human intervention and avoiding the inefficiency of per-sample generation. We introduce a semantic exploration algorithm named SEVEX, which addresses two major challenges of visual prompt exploration: (1) the distraction caused by lengthy, low-level code and (2) the vast, unstructured search space of visual prompts. Specifically, our method leverages an abstract idea space as a search space, a novelty-guided selection algorithm, and a semantic feedback-driven ideation process to efficiently explore diverse visual prompts based on empirical results. We evaluate SEVEX on the BlindTest and BLINK benchmarks, which are specifically designed to assess LVLM perception. Experimental results demonstrate that SEVEX significantly outperforms baseline methods in task accuracy, inference efficiency, exploration efficiency, and exploration stability. Notably, our framework discovers sophisticated and counterintuitive visual strategies that go beyond conventional tool usage, offering a new paradigm for enhancing LVLM perception through automated, task-wise visual prompts.

Low-resolution face recognition (LR-FR) remains a challeng-ing task due to poor feature extraction and aggregation, as probe imagesoften contain limited identity information resulting from extreme degrada-tions such as blur, occlusion, and low contrast. Additionally, the domaingap between high-resolution (HR) gallery images and low-resolution (LR)probe images poses a significant challenge. A single feature encoder strug-gles to generalize effectively across both domains when fine-tuned on anLR dataset, and this issue is further magnified by catastrophic forget-ting. To address these challenges, we propose FaceMoE, an effectiveadaptation of Mixture of Experts (MoE) transfomer architecture for low-resolution face-recognition . Specifically, we introduce multiple specializedfeed-forward network (FFN) experts and incorporate a top-k router, whichdynamically assigns tokens to appropriate experts. This design emergentlypromotes specialization across experts for different semantic regions ofthe face, which enables FaceMoE to perform resolution-aware featureextraction. Moreover, the top-k router facilitates sparse expert activation,enabling the model to preserve pretrained knowledge when finetuned on aLR dataset, while increasing model capacity without proportional compu-tational overhead. FaceMoE is trained with a combined face recognitionloss, router z-loss, and load balancing loss to ensure expert specializa-tion and stable training. To the best of our knowledge, this is the firstwork leveraging MoE for LR-FR. Extensive experiments across elevendatasets, spanning HR, mixed-quality, and LR benchmarks, demonstratethat FaceMoE significantly outperforms state-of-the-art methods.

The classification accuracy of pretrained Vision-Language Models (VLMs) relies on the quality of the text prompts. Handcrafted templates and Large Language Model (LLM)-generated descriptions not only make predictions more interpretable, but also enable reuse of the same prompts across heterogeneous VLMs. Recent works construct taskadapted text prompts with a small number of labeled images. However, existing few-shot text prompting methods do not explicitly focus on misclassified examples during prompt construction, leading to only marginal improvements even as more shots become available. To fully exploit few-shot supervision, we propose Text Prompt Boosting (TPB), an AdaBoost-inspired framework that treats each text-prompt-based classifier as a weak learner and sequentially aggregates them into a strong ensemble by explicitly targeting hard, misclassified examples. Extensive experiments show that TPB preserves task-intrinsic, model-agnostic cues in text space, enabling robust cross-model transfer. Across eleven classification benchmarks, TPB improves accuracy on the source model and preserves shot-driven gains when transferred to larger, more capable VLMs, where existing methods struggle to sustain such improvements.

Feed-forward 3D Gaussian Splatting (3DGS) has emergedas a highly effective solution for novel view synthesis. Existing meth-ods predominantly rely on a pixel-aligned Gaussian prediction paradigm,where each 2D pixel is mapped to a 3D Gaussian. We rethink this widelyadopted formulation and identify several inherent limitations: it rendersthe reconstructed 3D models heavily dependent on the number of inputviews, leads to view-biased density distributions, and introduces align-ment errors, particularly when source views contain occlusions or lowtexture. To address these challenges, we introduce VolSplat, a new multi-view feed-forward paradigm that replaces pixel alignment with voxel-aligned Gaussians. By directly predicting Gaussians from a predicted3D voxel grid, it overcomes pixel alignment’s reliance on error-prone 2Dfeature matching, ensuring robust multi-view consistency. Furthermore,it enables adaptive control over density based on 3D scene complex-ity, yielding more faithful Gaussians, improved geometric consistency,and enhanced novel-view rendering quality. Experiments on widely usedbenchmarks demonstrate that VolSplat achieves state-of-the-art perfor-mance, while producing more plausible and view-consistent results.

Text-to-image diffusion models are increasingly deployed increative contexts, yet remain impersonal, optimized for aggregate aes-thetics rather than individual taste. Human preferences are inherentlypluralistic: one user who favors muted, nostalgic portraits may prefer vi-brant compositions for street photography, while another gravitates to-ward dreamy, overexposed film aesthetics. Existing personalization meth-ods require dense interaction histories or per-user fine-tuning, failing incold-start settings and collapsing each user’s context-dependent prefer-ences into a single static representation. We introduce zero-shot imagepersonalization from personas (ZIPP), a paradigm that conditions im-age generation on natural-language personas (concise descriptors of auser’s identity, interests, and aesthetic sensibilities) without any user-specific data or weight updates. ZIPP uses an LLM in a roleplay set-ting to rewrite input prompts from the perspective of a given persona,steering diffusion models toward personalized outputs. To mine per-sonas at scale, we develop an inductive Graph Attention Network over a23M-user Reddit interaction graph with dual contrastive objectives thatalign graph structure with users’ visual behavior, and verbalize learnedrepresentations into coherent natural-language personas via an MLLM.We further introduce ZIP-Bench the first zero-shot image personal-ization benchmark, pairing 1.5K users with graph-mined personas and40K generated images. Across four benchmarks and 14 LLMs spanningfive model families, persona conditioning yields consistent improvementsin both zero-shot and few-shot settings, with frontier models achievingthe strongest gains (13–20%). In the few-shot setting, ZIPP matchesor exceeds fine-tuned baselines requiring per-user adapters trained on100+ examples. Unlike baselines that collapse preferences into a fixedstyle, ZIPP preserves intra-user preference diversity, achieving the low-est distributional divergence from users’ true preference distributions(CMMD 0.16 vs. 0.55 for fine-tuned alternatives). IPF-normalized eval-uation against global population demographics further reveals that ex-isting methods exhibit substantial bias toward narrow subpopulations,which persona conditioning significantly mitigates. A human evaluationconfirms these findings: ZIPP achieves a 79% win rate over genericgeneration and outperforms all fine-tuned baselines (58–65% win rate)without any user-specific training. We release ZIP-Bench at: https://huggingface.co/datasets/behavior-in-the-wild/zip-bench

We develop a method for producing vector sketches one partat a time. To do this, we train a multi-modal language model-based agentusing a novel multi-turn process-reward reinforcement learning follow-ing supervised fine-tuning. Our approach is enabled by a new datasetwe call ControlSketch-Part, containing rich part-level annotations forsketches, obtained using a novel, generic automatic annotation pipelinethat segments vector sketches into semantic parts and assigns paths toparts with a structured multi-stage labeling process. Our results indicatethat incorporating structured part-level data and providing agent withthe visual feedback through the process enables interpretable, control-lable, and locally editable text-to-vector sketch generation. Project page:https://xiaodan.io/teaching-an-agent-to-sketch/

Conventional vision-language models are largely object-centric, focusing on detecting and describing individual entities. In safety-critical X-ray baggage screening, however, threat often emerges not from a single object but from the functional compatibility of spatially dispersed components, such as batteries, detonators, and explosive charges. We formalize this setting as compositional threat reasoning, where risk is modeled as a relational property of grounded regions rather than an independent detection outcome. We introduce Falcon, a multimodal framework that abstracts segmentation-aware region features into a structured safety state capturing component presence, pairwise functional compatibility, and scene-level risk. This structured representation is injected into the language model as an explicit intermediate interface, encouraging relationally consistent and safety-aware reasoning. To evaluate this problem, we present Falcon-X, a benchmark that unifies dense grounding with structured supervision over component completeness and risk inference in cluttered X-ray imagery. Experiments show that while existing multimodal models adapt to appearance, they struggle with compositional safety reasoning. Falcon improves functional grounding and produces more coherent threat assessments, establishing compositional safety reasoning as a distinct evaluation paradigm for multimodal systems.