New Computer Vision (cs.CV) papers, Thursday 22 October 2026
138 papers announced on arXiv. Get them ranked for you, free →
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
AI summaryContinued pre‑training on 120 settings of model size, data volume, and world diversity shows that more data boosts in‑corpus performance, diversity improves transfer to related worlds, and larger models are more data‑efficient but generalize worse to distant worlds.
Different Defects, Different Effects: A Controlled Audit of Face Recognition Training Data
AI summaryThe paper audits MS1MV3, identifying five defect types (duplicates, split identities, crop errors, multiple‑face ambiguity, invalid content). Removing duplicates, merging split identities, and deleting invalid images improve IResNet‑50/ArcFace performance and training efficiency.
Motion Equivariant Convolutional LSTM: When Motion Becomes a Gauge Symmetry
AI summaryMotion Equivariant Convolutional LSTM (MELSTM) models sequences as gauge fields, making the network equivariant to any time‑varying motion. It outperforms standard CLSTM and flow‑equivariant CLSTM on future‑frame prediction, motion‑based digit classification, and action recognition, generalizing to unseen motions with fewer parameters.
Complexity-Aware Surgical Video Reasoning Segmentation with Digital Twin and Context Memory
AI summarySurgReasonSeg uses a digital‑twin foundation model with two LoRA‑enhanced language‑model adapters and an RL‑trained router to reason over surgical videos, adapting temporal windows and context memory per query complexity. It outperforms ReSurgSAM2 on EndoVis‑ReasonSeg by up to 0.297 J&F, running at 18.8 FPS.
JUICE: Unified Multimodal Modeling in Pretrained Representation Spaces
AI summaryJUICE trains a single flow‑matching Transformer on frozen image‑and‑text encoder spaces, jointly denoising both modalities. It achieves competitive image generation (48.41 GenEval) and VQA (73.1), while coupled sampling improves multi‑modal tasks like maze planning to 93.7% success.
SAIL: A Learnable Generalisation of Spatial Autocorrelation over Graph-Structured Data
AI summarySAIL learns a differentiable spatial autocorrelation operator for graph-structured data by jointly optimizing feature projections and neighbourhood affinities. It recovers classic Moran and Geary statistics and achieves Bayes‑optimal prediction for tasks that depend on latent spatial coherence.
Register Anything Model for Generalizable and Robust Point Cloud Registration
AI summaryRegister Anything Model (RAM) learns a single point‑cloud registration network that generalizes across indoor, outdoor, and object domains by geometry‑aware voxelization and a local‑consistency transformer, achieving state‑of‑the‑art inlier ratio and recall on 12 benchmarks.
WorldStream: Streaming Interactive 3D Worlds with Hierarchical Planning and Persistent State
AI summaryWorldStream extends 3D terrain on the fly, handles open-ended language commands, and preserves edits across revisits. It uses hierarchical planning to generate and compile macro‑level settlement/transport plans into micro‑tile details, enabling persistent, budget‑bounded interactive worlds.
HorizonComposer: Spatiotemporally Consistent Driving Video Editing with Enriched Traffic Semantics
AI summaryHorizonComposer edits driving videos while preserving traffic semantics by training on a hybrid real‑synthetic dataset with pixel‑level dense rewards and a “thinking‑frames” transition to avoid temporal shock. It achieves 39 % better content preservation, 12 % more temporal consistency, and 5 %–7 % improvements in downstream vision tasks.
Conditional Imaging for Privacy-Preserving Visual Sensing
AI summaryConditional imaging hides a scene in each raw capture by varying the camera’s optical state per frame; only with knowledge of that state can the image be reconstructed. Experiments on face data show authorized users recover identity, while unauthorized observers see only obscured images.
Sketch2Sim: an end-to-end modular framework for hand-drawn circuit analysis
AI summarySketch2Sim converts hand‑drawn RLC circuit photos into SPICE‑compatible netlists using YOLO component/text detectors, CRNN OCR, and DBSCAN node clustering, then solves for voltages and currents. It achieves robust accuracy across varied handwriting, lighting, and clutter.
Beyond Correct Answers: How Compositional Cues Shape Visual Judgments
Visual question answering requires models to assign attributes and relations to the correct objects, a problem known as visual binding. Structured prompts make these assignments explicit through labeled fields. Yet an improvement from the full prompt does not reveal which labels…
Depth-guided Geometric Adaptation for Few-shot Image Classification
AI summaryDepth-guided Geometric Adaptation (DAFIC) uses monocular depth to learn geometry-aware representations, refine patch features, and perform bidirectional dense matching, achieving state‑of‑the‑art few‑shot image classification on four benchmarks.
DualHeads: Reward-tilted few-step diffusion made simple
AI summaryDualHeads attaches two lightweight binary heads to a frozen few‑step diffusion model: one scores samples relative to the same prompt, the other distinguishes references from generated images. Optimizing the weighted sum of logits yields task rewards nearly matching 40‑step RL methods while keeping image quality and inference cost unchanged.
Real2Sem2Real: 3D Semantic Augmentations for Robot Learning from a Single Human RGB Demonstration
Human video scales easily and does not need robot hardware, but existing methods require depth sensing, overlay the robot onto 2D video, or generate data only for the demonstrated objects. We present Real2Sem2Real (R2Sem2R), a real-to-semantic-to-real pipeline that turns one…
Policy-Mask: Budget-Aware Adaptation of Diffusion Features to Discriminative Tasks
AI summaryPolicy‑Mask learns a budget‑aware policy to jointly pick timestep, block, and channel subset from a frozen diffusion model, normalizing scores to stabilize training. It outperforms fixed extraction baselines, achieving 38.9 mIoU on ADE20K and 75.4% top‑1 on CIFAR‑100 with only 128 channels.
Data-Driven Glyph Positioning for Handwritten Music Synthesis
AI summaryThe paper replaces hand‑crafted glyph placement rules with empirically derived constants and a graph neural network that predicts context‑dependent offsets, reducing placement error by up to 85 % and boosting downstream OMR accuracy by 30 % mAP50.
MINER: Modeling Inter- and Intra-Person Variance for Reidentification
AI summaryMINER suppresses garment bias in clothes-changing ReID by contrasting hard examples and adaptively balancing intra- and inter-person variance, achieving state‑of‑the‑art retrieval on four CC‑ReID datasets and top ranking in all tracks of the TVRID RGB‑Depth challenge.
Pair-Free Coarse-to-Real Video Generation via Real-to-Coarse Image Supervision
AI summaryThe paper trains a coarse‑to‑real video model without (coarse, realistic) video pairs, using only paired real‑to‑coarse images to generate pseudo‑coarse videos, then learns C2R mapping. It achieves scalable photorealistic video synthesis from any real‑video corpus.
RoPT: Robust 6DoF Object Pose Tracking in Video
AI summaryRoPT is a closed‑loop 6‑DoF tracking system that fuses mask‑based visual perception, a kinematic factor graph for regularized updates and multi‑hypothesis verification, and an active re‑localization module. It achieves higher survivability than open‑loop trackers on rapid‑motion benchmarks.
Can Quantum Layers offer Performance and Security Advantages in Vision-Language Models?
As quantum computing moves from the noisy intermediate-scale quantum (NISQ) era toward early fault-tolerant quantum computing (EFTQC), quantum-enhanced AI is shifting from a conceptual idea to experimental research. Quantum components are expected to extend the capabilities of…
TEMPL: A Resource-Efficient and Transparent Model-Free Framework for Pseudo-Label Generation in Material CT Scans
AI summaryTEMPL generates pseudo‑labels for material CT scans using thresholding, morphology, and breadth‑first search without pre‑trained models or GPUs. It achieves 63.19 % IoU and 33.3 % instance recall, outperforming zero‑shot SAM while running 15 s per image on a CPU.
MedBOE: Optimization-Based Estimation of Complete Patient Body Models from Whole-Body Medical Imaging
AI summaryMedBOE automatically reconstructs full patient body meshes from incomplete whole‑body scans by fitting a SMPL template using attribute‑guided optimization and constrained deformation. It outperforms existing methods on PET‑CT and MRI datasets, achieving <2% error in body‑habitus metrics.
T-DIG: Text-Conditioned Diversity for Image Generation
AI summaryT‑DIG adds user‑controlled diversity to any frozen image generator by using an LLM to produce axis‑specific prompts, trained with GRPO to maximize text‑conditioned diversity while maintaining prompt alignment. It outperforms baselines and works on open, distilled, and proprietary models.
Enhancing Cross-Grade Generalization in Banana Variety Classification via Hybrid CNN-Transformers and Knowledge Self-Distillation
AI summaryHybrid CNN‑Transformer architecture with KL‑divergence self‑distillation improves cross‑grade banana variety classification; achieves 81.25% mean accuracy on 31,678‑image dataset, outperforming pure CNN and ViT baselines, and enhances feature transferability and attention localization.
Single-Image Food Mass Estimators Do Not Transfer Their Gram Scale Across Datasets
AI summaryFine‑tuned DINOv2 on Nutrition5k dishes predicts food mass with 24.8 g MAE on that dataset, but fails to generalize to external datasets, where zero‑shot Qwen3‑VL outperforms it. Accurate estimation requires labeled images from the target setting.
HumanCentric-Scenes-AI: A Human-Curated Multimodal Dataset of AI-Generated Human-Centric Scenes Across Surveillance, Occupational Portraits, AI-Edited Images, and Short-Form Video
AI summaryHumanCentric‑Scenes‑AI is a 296‑image/video dataset of AI‑generated human scenes, split into surveillance, occupational portraits, edited photos, and short videos, with prompt, metadata, and quality reports. It enables perception studies, AI‑detection calibration, and bias analysis.
Weakly Supervised Fish Segmentation through Point-Supervised Tracking and Motion-Informed Language
AI summaryWeakly supervised fish segmentation using only one click per fish. The method predicts tracking points, converts them to motion-informed language cues, and feeds them to a language-conditioned segmentation model, achieving higher mIoU on DeepFish with minimal labeling.
OccJEPA: Occlusion as Masking for Amodal Video Segmentation
AI summaryOccJEPA treats occlusion as masking, using a pre‑trained V‑JEPA 2.1 predictor to fill in unseen pixels and a light head to decode amodal masks. It is deterministic, 33.5× faster than diffusion methods, and outperforms them on SAIL‑VOS, TAO‑Amodal, and Kubric.
ReSeg: Retraining-free segmentation of abdominal organs via generalizable semantic representation learning
AI summaryReSeg trains a generalizable semantic representation network on synthetic data and then fine‑tunes an encoder‑decoder for each organ, achieving zero‑shot abdominal segmentation. It outperforms SOTA across CT, MRI, and PET datasets without using any real images.
UNIVERS: Learning UNIversal Voxel-wise Embedding via Randomized Synthesis for generalizable one-shot landmark localization
AI summaryThe paper introduces UNIVERS, an end-to-end framework for one-shot landmark localization that learns universal voxel-wise embeddings using randomized shape and appearance synthesis. Experiments on multi-modality adrenal gland datasets show that UNIVERS outperforms existing state-of-the-art methods, especially in cross-modality scenarios.
Diffusion Posterior Sampling with Uncertain Forward Models
AI summaryDiffusion posterior samplers can incorporate parameter uncertainty in the forward model via an Uncertainty-Aware Likelihood (UAL), a Gaussian approximation to the marginal likelihood. Using UAL improves reconstruction quality by 9–92 % MSE and up to 6 dB PSNR across multiple inverse problems.
From Canon to Nikon: Camera Domain Shift and Potential Equity Implications in Retinal AI
AI summaryThe study shows Nikon retinal cameras produce poorer quality images than Canon, leading to more AI misclassifications; performance differences vary by task, resolution, and classifier, indicating camera choice is a major, often overlooked source of AI variability in retinal diagnostics.
DriftNeRF: Cross-Ray Density Drifting for Sparse-View Neural Radiance Fields
Neural Radiance Fields (NeRF) synthesize photorealistic novel views from dense image collections, but degrade sharply when few views are available. We attribute a specific, measurable defect to the objective itself: the rendering loss is a sum over independent rays and contains…
PhyMo: Learning Physical Dynamics with Accurate and Continuous Motion from Multi-View Videos
AI summaryPhyMo learns accurate, continuous 3D motion from multi‑view videos by using a Motion‑Aware Trajectory Representation to avoid appearance shortcuts, and a bidirectional kinematic continuity constraint for temporally consistent displacement, velocity, and acceleration, achieving state‑of‑the‑art reconstruction and extrapolation.
WaFIR: Wavelet-Fourier Mixture-of-Experts with Subband-Adaptive Routing for Blind All-in-One Image Restoration
AI summaryWaFIR decomposes image features into Haar subbands, summarizes each with a Fourier descriptor, and routes them to prompt‑conditioned spectral experts. It boosts average PSNR by 0.48 dB over PromptIR on denoising, deraining, and dehazing, without extra parameters.
Two for One: Privileged Learning for Resource-Constrained Screening
AI summaryThe paper proposes a privileged learning method where a teacher trained on bilateral CFP images transfers knowledge to a student that only needs one eye at test time. The student outperforms conventional monocular models, improving AUROC from 0.931 to 0.966 for AMD and 0.789 to 0.836 for DR.
DGCM-SSFA: Dual-Granularity Cross-Modulation and Sparse Strip Frequency Attention for Efficient 3T-to-7T MRI Synthesis
AI summaryDGCM-SSFA synthesizes 7T MRI from 3T scans with only 11M parameters, using dual-granularity cross-modulation and sparse strip frequency attention to preserve edges and global context. It achieves SSIM 0.8431 (T2w) and 0.8100 (T1w), rivaling larger Transformers.
From Augmentation to Synthetic Evidence: Evaluating Generated Chest Radiographs at Dataset Intake
AI summaryThe paper evaluates how well automated checks can detect synthetic chest X‑ray images and finds that while quality filtering can improve acceptance rates, no strategy reliably ensures both quality and authenticity. It concludes that source provenance must be recorded separately for dataset inclusion.
SICAF: Time-Varying Focus Bottleneck for Self-Supervised Event-Based Optical Flow with Spiking Neural Network
AI summarySICAF trains self-supervised spiking neural networks for event-based optical flow by compressing latent features while applying temporal attention to preserve saliency. It outperforms baselines, reducing average end-point error by 8.14% on MVSEC and boosting robustness to adversarial attacks by 13.49%.
RiO-DETRv2: Geometry-Aware Uncertainty Modeling Meets Real-Time Oriented Object Detection Transformer
AI summaryRiO-DETRv2 introduces geometry‑aware uncertainty modeling for real‑time oriented object detection, using object‑aligned boundary distributions, circular‑to‑local angular residuals, and unified OBB quality alignment. It improves DOTA mAP50‑95 by 1.2–1.5 points over RiO-DETR with unchanged speed.
Carving Radiology Reports at Their Evidential Joints: Anatomical Region Sets as Generation Units
AI summaryRadCarve generates radiology reports by carving them into units defined by anatomical region sets extracted from statement-region correspondences. It jointly predicts unit existence and region composition, yielding a 9.4% CXRFEScore boost, 2.4× higher binding F1, and state‑of‑the‑art report results with far fewer parameters.
How Representational Priors Shape Generalization in World Action Models
World action models (WAMs) are shown to substantially diverge under generalization performance based on their generative backbones. We investigate this gap through the representational geometry that WAMs inherit from their backbones and how policy training reorganizes it. Our…
LocalAgent: Collaborative Agentic Verification for Fine-Grained Instance-Level Consistency
AI summaryLocalAgent combines a general MLLM with lightweight Local Evidence Verifiers that analyze suspicious regions for appearance and geometric consistency, using reinforcement fine‑tuning to trigger verification only when needed. It improves mean accuracy by 4.9% over the best baseline while invoking the verifier on only 12.13% of cases.
Multi-Scale Latent Temporal Compression Attention and Video-Text Cross Attention Fusion for Video Action Detection
AI summaryThe paper introduces a Multi‑Scale Latent Temporal Compression Attention module to aggregate temporal information at varied scales and a Video‑Text Cross Attention Fusion module for fine‑grained alignment of visual and textual features. It reports state‑of‑the‑art action detection accuracy on multiple benchmarks.
ROCOCO: Robust Outline Completion of Occluded Curvilinear Objects
AI summaryROCOCO uses flow-matching generative models to inpaint occluded curvilinear objects without a prior mask, augmented with a UNet that predicts the mask from the observation. It outperforms baseline inpainting methods, achieving state‑of‑the‑art full‑image error with far fewer parameters.
A Unified Evaluation Framework of Physical Adversarial Robustness in the Edge VLM Era
AI summaryThe paper introduces a unified metric, conditional attack success rate (cASR), to evaluate physical adversarial patches on edge vision‑language models and object detectors. It finds VLMs are as vulnerable as detectors, with failure modes shifting from omission to displacement/substitution under distribution shift.
NeuralSRender: G-Buffer Guided Video Diffusion for Sim-to-Real Rendering
AI summaryNeuralSRender translates synthetic driving videos into realistic ones by conditioning a video diffusion transformer with G‑buffer priors, using a lightweight projector and LoRA for efficient adaptation. It improves detection AP from 35.1 to 40.6 and segmentation mIoU from 32.4 to 40.8 while adding only 1.45% more parameters.
DetM-Tracker: Deterministic and Memory-Efficient VIS for Autonomous Industrial Inspection
Autonomous industrial inspection requires strictly deterministic real-time inference under tight memory and computational constraints. However, existing Video Instance Segmentation (VIS) models suffer from prohibitive scaling: dense temporal memory grows with spatial resolution…
Cache Locally, Recompute Selectively for Streaming Video Understanding
Streaming video understanding (SVU) requires video-LLMs to process incoming video continuously while answering user queries with low latency. Recent methods reduce query-time cost by prefilling frames into KV caches during idle time and reusing the cached KVs at query time. To…
DisasterRoute: Evidence-Guided Dual-System Routing for Visual Disaster Question Answering
AI summaryDisaster-S1 is a vision‑language model that produces an answer in a single forward pass, achieving fast, accurate predictions. DisasterRoute routes queries to Disaster-S1 or the more complex DisasterVL based on confidence and evidence, boosting accuracy while keeping latency low.
Think-Capture: Active 3D Geometry Acquisition for Spatial Reasoning in Vision-Language Models
Vision-Language Models (VLMs) excel at 2D recognition but often fail to maintain consistent 3D relations across sparse viewpoints, leading to spatial hallucinations under occlusion and perspective changes. Existing methods commonly fuse all frames with dense geometry, incurring…
Encoded but Unused: Bounding the Sources of VLM Errors
AI summaryThe paper introduces VELens, a rank-based probe that estimates how often vision-language models fail to use encoded visual evidence. It finds that visual-utilization errors account for 84.4% of LLaVA mistakes and 70% of Qwen mistakes, highlighting a key bottleneck.
Failure Modes of Vision-Language Models for Human Activity Understanding in Thermal HRI
AI summaryThe study evaluates seven vision‑language models on thermal camera data for human activity understanding in HRI, finding they can classify broad activities but fail at fine‑grained actions, mis‑time events, hallucinate objects, and show inconsistent abstention, limiting robotic reliability.
DepState-JEPA: Learning Depth-Grounded State Representation for Visual Perception and Manipulation
Predictive representations for robot manipulation should connect scene geometry with the state changes induced by actions. Our intuition is that understanding what an action will change starts with understanding where things are: complementary RGB and depth observations ground a…
ER-Chord: Spatially Consistent One-Step Image Editing via Equal-Energy Reliability Redistribution
AI summaryER‑Chord redistributes one‑step diffusion editing strength using local reliability inferred from agreement across noise levels, preserving total control energy without extra denoiser calls. It improves source preservation and spatial consistency over state‑of‑the‑art baselines on PIE‑Bench and TEdBench.
CRAB: Cuboid-based Representation for As-built BIM from Point Clouds
AI summaryCRAB reconstructs as-built BIM from point clouds using cuboid representations for walls, doors, windows, beams, and columns; it achieves 2.80 cm mean wall thickness error and outperforms plane-based methods in completeness, accuracy, and generalization.
Clinical radar heart sounds with echocardiographic phenotypes of ventricular dysfunction and valvular disease
AI summaryNon‑contact radar recordings of 537 patients were paired with echocardiographic reports for 162 cases. The dataset enables analysis of cardiac mechanical signals, achieving an AUROC of 0.770±0.138 for detecting reduced ejection fraction.
Certified Abstention for Provably-Unidentifiable Index-Phase Recovery in Image Registration
AI summaryThe paper proves that in reference‑perturbed imaging, position and phase recoveries are independent, derives a tight resolution threshold separating recoverable from provably‑unidentifiable registration, and introduces a certified‑abstention framework that exactly recovers identifiable regions while abstaining on unidentifiable ones.
Computer Vision and Knowledge-Grounded Language Models for Intelligent Beekeeper Assistance
AI summaryWe build an edge‑deployable pipeline that detects comb‑cell classes from hive photos, converts detections to a structured fact schema, generates constrained factual text, and uses a knowledge base to provide actionable beekeeper recommendations. We release FAIRHiveFrames‑1K and FAIRHiveText‑1K datasets.
Latent space configuration for neural network finetuning, fast ranking, and embedding dimensionality reduction
AI summaryLatent Space Configuration finetunes ArcFace‑IResNet50 embeddings to give a structured latent space, enabling O(1) candidate selection. This yields up to 3× GPU and 40% CPU speed‑ups over exhaustive search while preserving or improving rank‑N accuracy, even with dimensionality reduction.
Asking What a Director Sees Can Change What Vision-Language Models Choose
Vision-language models are now routinely tested on theory of mind: whether they can work out what another person sees, knows, and means. A common design asks the model to report that perspective and then to act for the person, and reads a correct report beside a wrong action as…
OpenMAP-HIE: Pathology Informed Score-based Diffusion Model for Unsupervised Neonatal Hypoxic-Ischemic Encephalopathy Lesion Segmentation
Accurate localization of neonatal hypoxic--ischemic encephalopathy (HIE) lesions on apparent diffusion coefficient (ADC) images is clinically important but is limited by scarce expert annotations and substantial variation in lesion appearance. We propose OpenMAP-HIE, a…
OACap: Caption-Guided Active Observation for Detailed Video Captioning
AI summaryOACap is a caption-focused agent that requests targeted observations to fill missing info, then revises captions iteratively. It improves video‑captioning metrics and introduces FSCap‑Bench with CSFS and CSRS scores to evaluate factual and semantic fidelity.
Reading the Mask: Descriptor-Grounded BI-RADS Rationale Generation for Breast MRI Segmentation
Deep learning based models accurately delineate tumours on dynamic contrast-enhanced MRI (DCE-MRI), but segmentation masks alone do not provide the rationales or reasoning used by radiologists to assess the lesion suspicion. This limits the explainability as well the clinical…
Seeing through Denoising: Frequency-guided One-step Denoising for Infrared Small Target Detection
Infrared small target detection (IRSTD) is a challenging task due to the extremely small size of targets, low signal-to-noise ratios, and heavy background clutter in infrared imagery. While the progress of deep learning has advanced this field, existing methods primarily adopt…
AOV-GS: Active Open-Vocabulary Semantic Mapping with 3D Gaussian Splatting
Robots deployed in human-centered environments need maps that support both geometric navigation and open-vocabulary semantic queries beyond a fixed label set. A key challenge is enriching such maps with semantic information without relying on manually annotated segmentation…
AdvFidelity: Unifying Preference Alignment and Few-Step Image Generation within Reinforcement Learning
Practical image generation requires both human preference alignment and few-step efficiency, yet combining reinforcement learning (RL) with distillation is often multi-stage and may erode RL gains We observe that few-step RL progressively improves structural coherence, even with…
FM-Occ: Feature-Diverse Upsampling and Memory-Aware Decoding for 3D Occupancy Prediction
3D semantic occupancy prediction estimates voxel-wise occupancy and semantic labels throughout a scene. Our experiments reveal that conventional sparse transposed-convolution upsampling yields poorly differentiated voxel responses and that successive decoder layers discard…
Progressive Permutation Alignment with Temperature-Controlled Spectral Sharpening
In implicit neural representations (INRs), 3D signals are encoded directly in network weights, making stable weight-space operations essential. However, permutation symmetry causes independently trained yet functionally equivalent models to appear distant in weight space, often…
Region-Aware Multi-View Inspection Using Deep Learning and Domain-Informed Decision Fusion
Automated optical inspection of precision components with tiny surface defects remains challenging when different regions of the same component exhibit different defect characteristics and, therefore, different acceptance criteria. This paper presents a region-aware, dual-model…
VideoGen Model 1: Prompt Expansion Versus Formatting in Audiovisual Generation
Creative prompt expansion improved overall audiovisual quality scores over raw prompts in a controlled pilot of VideoGen Model 1, our post-trained version of MiniMax H3. We compared the raw, reformatted, and expanded versions of 30 prompts, holding the generator, seed, and…
LiveAct: Towards Real-Time Multi-Agent World Models with Heterogeneous Action Interfaces
AI summaryLiveAct trains a real‑time multi‑agent video world model that interprets heterogeneous controls by mapping each action stream to a visual agent via language‑guided spatial gates and schema‑guided semantic adaptation. It attains 0.898–0.942 attribution accuracy and 17.7 FPS on six game and robot tasks.
LLMSeg: Training-Free Text-Prompted Image Segmentation with B-Splines
AI summaryLLMSeg turns a vision‑capable LLM into a training‑free, text‑prompted image segmenter by iteratively localizing instances, tracing their borders with B‑splines, and refining until masks stabilize. It matches or outperforms SAM on ReasonSeg, RefCOCO, and LVIS without any segmentation training.
CCG Diffusion: Combinatory Categorial Grammar-Driven Compositional and Syntactically Disambiguated Text-to-Image Generation
AI summaryCCG Diffusion uses Combinatory Categorial Grammar to read syntactic relations directly from a parse, then rewrites prompts and guides attention in diffusion models. It achieves top performance on attribute binding, spatial relations, and prepositional phrase disambiguation for Stable Diffusion and Sana.
Look Beyond the Question: Preserve-Focus-Fuse for Whole-Slide Image Question Answering
A question about a whole-slide image rarely names every morphological finding needed to answer it. Question-guided patch selection can therefore overlook supporting morphology, while slide-level compression can dilute focal evidence. Inspired by how pathologists combine broad…
Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers
Full 3D attention is costly in video Diffusion Transformers. Block-sparse attention reduces this cost by computing only selected attention blocks. However, existing methods either overlook the potential to optimize sparsity allocation or rely on heuristic rules or local metrics…
Beyond Shortcut Learning: Hierarchical Spatial Prototype Regularization for Face Forgery Detection
Recent advances in generative models have enabled highly realistic face forgeries, yet generalizing deepfake detectors to unseen manipulation pipelines remains a major challenge. One critical yet underexplored issue is that the learned forensic representation itself may fail to…
BlazeVGGT: 13x Faster VGGT at 2,000 Frames via Camera Token without Training
Feed-forward visual geometry models, exemplified by VGGT, have demonstrated strong multi-view reconstruction capabilities through dense global attention. However, the resulting all-to-all cross-view computation scales quadratically with the number of input views, making…
Extending Affordance Segmentation for Wearable Robots to Multi-Object Images
AI summaryExtends vision‑based affordance segmentation for wearable robotic prostheses to multi‑object scenes by adding a laser pointer for disambiguation and retraining CNNs with procedurally mixed‑reality data. Improves disambiguation accuracy by up to 9% without extra compute.
Balancing Foundation Models and Trait-Specific Learning for Forehead-Crease Recognition
Biometric recognition often degrades under partial facial vis- ibility, where conventional facial cues are occluded. The forehead region remains consistently visible and contains distinctive crease patterns that can serve as a complementary biometric trait. We present a unified…
AutoPhyX: Automatic Text-Condition Physics Property Generation
High-fidelity 4D physical simulation requires accurate physical parameter assignments that align with diverse material properties. Previous methods struggle with critical bottlenecks: they either rely on computationally intensive per-scene optimization or predict parameters from…
Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction
Curriculum learning orders samples by difficulty, but difficulty alone does not establish supervision reliability for dense prediction. Easy samples can expose shortcuts, hard samples can be informative, and noisy or ambiguous samples can mislead training. Building on our…
Melange: A Synthetic Visual Bias Benchmark for Multi-modal Large Language Models
Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal…
Hyperspherical Simplex Representations for Backward-Compatible Retrieval
Training modern AI models has become increasingly expensive, and updating deployed models can alter their internal feature representations, affecting applications built upon them. In retrieval systems, such updates often require complete gallery re-indexing, a computationally…
Generalized CycleGRPO for Fine-Grained Multimodal Understanding and Grounding
Fine-grained multimodal intelligence requires models to both understand a designated region and localize it from language. Existing methods learn these directions with separate objectives and task-specific annotations. We present Generalized CycleGRPO, the first unified…
Rethinking Underwater Image Enhancement: Learning through Text-Guided Degradation Curriculum
Existing paired clean-degraded datasets have advanced learning-based underwater image enhancement, but they remain either limited in scale or difficult to acquire, and still fail to adequately cover the diverse degradations encountered in real-world underwater environments. Such…
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the…
CineEvolve: An Agentic Self-Evolution Framework for Cinematic Multi-Shot Video Generation
Cinematic multi-shot video generation requires both narrative planning and iterative refinement. Existing methods either lack narrative reasoning or rely on feed-forward workflows without policy updates. In this paper, we propose CineEvolve, an agentic self-evolution framework…
No Frames, Just Frequencies: Bandwidth-Efficient Deployment-Time Adaptation for Smart Glasses
AI summarySmart glasses adapt to distribution shifts by sending tiny frequency descriptors instead of frames, enabling a host to synthesize out‑of‑distribution data and fine‑tune the model. The method restores 93–95% of supervised adaptation gains with 213×–784× bandwidth savings.
Diff-IrisPAD: Modeling Bona Fide Iris Distributions for Unsupervised Presentation Attack Detection
AI summaryDiff-IrisPAD models bona fide iris textures with a Brownian Bridge Diffusion network, reconstructing clean irises and flagging attacks by reconstruction inconsistency. It achieves a 30.52 ACER on 125K+ multi‑sensor images, outperforming diffusion baselines and generalizing well to unseen attacks.
An interpretable deep learning method for early hepatocellular carcinoma diagnosis on multiphase CT
Early diagnosis of hepatocellular carcinoma is critical to preserve access to curative treatment options. The Liver Imaging Reporting and Data System (LI-RADS) provides a standardized framework for non-invasive diagnosis on multiphase computed tomography, based on major visual…
TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation
Edge medical segmentation must reconcile small model size, high accuracy, and fast inference. TernScope addresses this tradeoff through model--hardware co-design, learning a compact segmentation network whose low-bit operators also simplify hardware execution. Split-channel…
FogTTA: Online Test-Time Adaptation for Robust Transformer-based Object Detection in Foggy Weather
AI summaryFogTTA adapts a Transformer-based detector online to foggy conditions by using the deployed model as a teacher to generate pseudo labels, refining them, and training a student with Varifocal loss; it outperforms prior methods while remaining real‑time.
OpenEarth-OVS44: A Multimodal Remote Sensing Dataset for Open-Vocabulary Segmentation
In remote sensing, open-vocabulary segmentation is particularly important for flexible and scalable land-cover understanding across diverse regions, as it improves generalization to novel categories without requiring exhaustive pixel-level annotations. Despite its great promise,…
Joint Semantic-Channel Coding and Modulation for Generative Image Transmission
Digital image transmission typically optimizes source coding, channel coding, and modulation separately, limiting how source representations adapt to channel impairments. We propose SemQAM, a joint semantic-channel coding and modulation scheme for generative image transmission…
Reading the Room: Scale-ambiguous Semantic Registration of Sparse VSLAM Maps to Floor Plans in Partially Built Environments
Many robotic construction applications, such as progress tracking, require registering a sensor trajectory to a prior 2D floor plan. Visual capture is attractive for its low cost and availability, but aligning a visual simultaneous localization and mapping (VSLAM) map to a floor…
MS2CAST: A Video Anomaly Detection Framework via Multi-Scale Spatial Cross-Attention and Efficient Spatiotemporal Networks
AI summaryMS2CAST combines a multi‑scale cross‑attention encoder (MViTv2‑S + ConvNeXt‑T) with interchangeable spatiotemporal modules (SwinLSTM or Vision Mamba) to predict future frames without optical flow, achieving competitive VAD AUCs on UCSD, CUHK, ShanghaiTech while using <1 GB GPU memory.
Spatially Gated Demographic Steering for Debiasing Multi-Person Text-to-Image Generation
AI summarySpatially Gated Steering independently steers each person’s demographics in multi‑person text‑to‑image generation by masking per‑person attention maps and applying distinct demographic vectors. It achieves high intra‑image gender and racial diversity while maintaining aggregate balance and image quality, without fine‑tuning.
Learning to Reason Like a Cytopathologist:Clinically Grounded Multiscale Evidential Learning for Cervical Cytology
Cervical cytology presents a fundamentally ambiguous diagnostic landscape, where equivocal categories such as Atypical Squamous Cells of Undetermined Significance (ASC-US) occupy a morphological grey zone that is definitional. For clinical decision support, the goal extends…
Rethinking Post-hoc Debiasing of Large Pretrained Models
Large pretrained models such as CLIP are widely used as frozen feature extractors. However, their embeddings inherit spurious correlations from training data, introducing bias into the features. A growing body of research has sought to address this issue through post hoc…
Efficient Generative Image Compression with Quantized Latents Refinement
Generative image compression (GIC) delivers strong perceptual quality at low bitrates. However, GIC methods with strong perceptual quality often rely on pretrained diffusion models with high latency, while recent lightweight GIC methods reduce latency but deliver lower…
Forward-Only Learning for Shared-Kernel Wavelet Neural Networks without Backpropagation
Backpropagation remains the dominant training paradigm, using backward gradients and local updates that can suffer from vanishing or exploding gradients in deep networks. Although the Forward-Forward algorithm avoids conventional backpropagation, it remains mainly local or…
Learning to Measure and Diagnose: A Clinical-Dictionary Radiomic Bottleneck Model for Thyroid Nodule Classification
Thyroid ultrasound supports the assessment of nodule malignancy, but classification predictions often lack evidence linked to familiar clinical descriptions. Radiomics provides explicitly defined measurements that can support such explanations through a clinical dictionary.…
APED-CXR: Adaptive Privileged Evidence Distillation for Chest X-Ray Grounding with Image-Only Inference
Compact chest X-ray (CXR) models that localize findings suit deployment, but box annotations are limited and report-aware multimodal models are expensive at inference. We introduce APED-CXR, which distills report-informed spatial evidence from a frozen multimodal teacher into a…
LEAF-Mamba: Feature Adaptation and Fusion for Whole Slide Image Analysis
AI summaryLEAF‑Mamba jointly adapts frozen patch features via a reconstruction‑regularized adapter and fuses them with parallel Mamba and convolutional streams, gated by input, to improve WSI diagnosis and survival predictions. It outperforms strong MIL baselines across feature extractors.
UniRAP: Towards Unified Part-level Physical Affordance Reasoning and Actionable Perception
Vision-language perception has achieved impressive progress in aligning natural language with visual observations, yet grounding high-level semantics into part-level physical interaction remains challenging. To address this gap, we propose UniRAP, a unified model for inferring…
Key Findings for Small UAS for Post-Disaster Damage Assessment from the CURSE'26 Exercise
The CLARKE building and road damage assessment AI classifier was incorporated into the Catastrophic Unmanned Remote Sensing Exercise (CURSE) 2026, which trained 23 small uncrewed aerial systems (sUAS) teams from state and local agencies and non-governmental organizations in…
JointLight: Human-Centric Video Relighting with Asynchronous RGB-Irradiance Joint Modeling
Understanding and modeling light transport in human-centric scenes is a longstanding challenge. While recent progress in diffusion models has boosted research in relighting, precise illumination control still lags, primarily due to the scarcity of diverse, high-quality,…
Open-Vocabulary Models and Label Propagation for Uncertainty-aware Pseudo-Labeling of Object Detection Datasets
Object detection models rely on large-scale annotated datasets that require substantial manual annotation effort. The labeling costs further increase when annotation uncertainty is captured through soft labels, as multiple annotations must be collected for each object. We…
Compositionality with Primitives: A Unified View from Euclidean to Hyperbolic Vision-Language Models
Compositional zero-shot learning (CZSL) aims to recognize novel combinations of previously observed objects and attributes. Prior work showed that large-scale vision-language models encode decomposable semantic structures that can be recomposed to enable generalization. The…
GuidedSAM-Plume: An Expert-Guided Segmentation Framework for Volcanic Plumes
Quantitative analysis of environmental phenomena increasingly relies on image-based measurements. In volcanology, the geometric characterization of volcanic plumes and clouds from RGB images remains hindered by the scarcity of dedicated segmentation solutions and partially…
Early versus Late RGB-Infrared Fusion for Counter-UAV Detection: A Comparative Study
This paper presents a novel strategy for UAV detection through visible (RGB) and infrared (IR) fusion, in order to improve accuracy and precision in harsh environments and weather conditions. In this research, two opposite approaches have been studied and implemented: the first…
Revealing Hidden Failure Modes in Chest X-ray Classification via Spectral Domain Analysis
Deep learning models for chest X-ray anomaly detection remain vulnerable to subtle distributional shifts (e.g., acquisition technique, patient-related factors, and preprocessing). Traditional error analysis often relies on semantic metadata or model embeddings, which can mask…
How Do DiTs Denoise? An Aggregator-Predictor View of Attention and MLPs
How do diffusion transformers (DiTs) represent image content and learn rules for reusing it to generalize? We develop an aggregator--predictor view of DiTs as input-adaptive local denoisers, in which context aggregation and patch prediction provide functional roles for…
GeoMem-VLA: Internalizing 3D Scene Understanding into Temporal Memory for Vision-Language-Action Models
Geometric structure is essential for robotic manipulation and understanding the 3D world. Most VLA models rely on 2D representations of current observations, while geometry-aware approaches often require additional 3D information. However, geometry inferred independently from…
Stochastic Nonlinearities Improve Uncertainty Estimation
Reliable uncertainty estimation is important across visual recognition tasks, providing information about the confidence and reliability of model predictions. Many uncertainty estimation methods obtain predictive diversity through mechanisms that can suppress intermediate…
Expandable Modality Experts for Medical Vision-Language Learning
Learning across medical imaging modalities requires making use of images without paired text and accommodating new domains while retaining existing representations. We propose a medical vision-language technique built around an expandable bank of modality experts. Its Multiway…
ViBMR: A Large-Scale Multi-Sequence Brain MRI Dataset and Comprehensive Evaluation for Vietnamese Radiology Report Generation
Automated radiology report generation for multi-sequence brain MRI holds strong promise for alleviating radiologists' workload, yet it remains a demanding problem: brain MRI studies are inherently volumetric, each pulse sequence carries complementary diagnostic information, and…
Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home
Thermal cameras can watch over older people living alone without capturing faces, and multimodal large language models (LLMs) can judge a short video without task-specific training. We evaluate a production multimodal LLM, Gemini 3.8 Flash at temperature 0, as the decision stage…
CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion
Diffusion-based video generation models (VGMs) can produce visually plausible videos, yet their camera ego-motion often drifts over long horizons. Camera ego-motion determines how the viewpoint changes between frames and is essential for maintaining consistent geometry and…
DewHand: Direct Metric World-Space Hand Motion Reconstruction from Egocentric Videos
Recovering metric world-space hand motion from egocentric video supports understanding human interactions and transferring them to embodied agents. Three challenges complicate this task: multi-stage pipelines propagate errors from independently estimated hand and camera motion;…
Existence of Minimizers for the Gaussian Mumford-Shah Model via Covariance Floors on the Value Distribution
The Gaussian Mumford-Shah model segments an image into regions by minimizing a continuous multivariate Gaussian negative log-likelihood term plus a perimeter penalty. It is used for scalar- and vector-valued images, in particular hyperspectral images, but the existence of a…
Training-free Acceleration for Blind Compressed Image Restoration via Activation-Calibrated Singular Value Decomposition
Transformer-based models achieve strong performance on blind compressed image restoration (CIR) under unknown compression settings, but their densely stacked projection operators make deployment costly. We study training-free acceleration for such models through an empirical,…
Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
AI summaryThis paper introduces Future Chunk Prediction, an auxiliary training objective for autoregressive video diffusion models that predicts temporally separated chunks without intermediate context. The method improves physical plausibility and visual evolution modeling while reducing the overall training compute budget.
DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
Despite recent progress in dexterous reference tracking, existing models focus primarily on single-hand manipulation and a limited range of interactions, leaving the bimanual dexterous behaviors common in real life unaddressed. The scarcity of contact-rich bimanual manipulation…
Multiclass Boundary Refinement with Neural Cellular Automata for Remote Sensing Land Cover Segmentation
Semantic segmentation of remote sensing imagery often places class boundaries a few pixels away from where they truly lie, and this position error is concentrated in the thin band measured by Boundary IoU (BIoU). We treat boundary refinement as the iterated application of a…
ClunoGS: Compact Language 3D Gaussian Splatting via Unified Optimization
AI summaryClunoGS compresses language‑aware 3D Gaussian splatting by jointly optimizing sparsification and quantization under constraints that balance rendering quality, semantic accuracy, and size. On LERF and ScanNet‑10, it preserves open‑vocabulary localization and segmentation far better than sequential or semantics‑agnostic compression.
RealMe: Immersive Volumetric Video from Minimal Captures
This paper addresses the problem of creating immersive volumetric videos from sparse and low-resolution capture resources. Conventional dense multi-view reconstruction methods can produce high-quality volumetric videos, but their reliance on costly capture setups makes them…
Plausible but Unimplantable: A Trustworthiness Audit of Shape Completion Foundation Models for Cranial Defect Reconstruction
A patient who has lost part of the skull vault needs an implant that fits the defect exactly, and such implants are still drawn by hand in CAD by specialists, which makes each case slow and expensive. The task can be considered as a shape completion problem, so pretrained…
Attn-Forcing: Learning Full-Context Attention Proxies for Streaming Video Generation
AI summaryAttn‑Forcing introduces learnable visual latents (“attention proxies”) that approximate full‑context attention in streaming video generation, using a Proxy Update Module and an Attention Matching Loss to enforce full‑context supervision. The method yields state‑of‑the‑art quality with bounded memory.
ProactEgoMem: Benchmarking Proactive Memory in Streaming Egocentric Video
Today's multimodal assistants are typically prompted to interpret the visual evidence presented to them. To offload the cognitive burden of remembering across long horizons, wearable assistants should instead proactively acquire and maintain useful memory as the user's…
AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
Vision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills. Reinforcement learning (RL) offers an effective avenue for further improving VLA policies. However, existing…
Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use
Multimodal agents (MMAs) sustain interaction through reasoning, memory, tools, and feedback, most visibly as computer-use agents, while robotic systems (RSs) couple sensing and actuation under physical dynamics. We define multimodal embodied agents (MMEAs) as goal-directed…
ALPS: Visual State Encoders for World Models
World models turn visual representations into predictions and actions. We argue that their visual foundation should combine physical-state information, native language alignment, and pixel reconstruction, so that a single representation can support prediction, instruction, and…
PhysiCalWorld: In-Context Physics Calibration for World Models
Existing world models typically perceive a scene through current visual frame or a narrow history window. However, these observations may not reveal the physics properties that determine transition dynamics, such as an object's mass distribution, joint damping, or surface…
VoxProg: 3D Spatial Reasoning through Voxel Programming
Language models' abilities to interpret tasks and generate programs offer a promising basis for 3D spatial reasoning. Yet following spatial instructions in 3D scenes requires programs to construct geometric references that language leaves implicit. In this work, we introduce…
Rethinking Video Autoencoders for Interactive World Models
Video world models predict future frames from past frames and actions. They serve as simulators in which agents train, plan or are evaluated, and that humans can control interactively. They generate in the latent space of a video autoencoder, whose compression sets their cost…