New Computer Vision (cs.CV) papers, Monday 19 October 2026

127 papers announced on arXiv. Get them ranked for you, free →

1. Revealing the Impact of Visual Text Style on Attribute-based Descriptions Produced by Large Visual Language Models

   Xiaomeng Wang, Martha Larson, Zhengyu Zhao · arXiv:2604.27553

   When the visual style of text is considered, a wide variety can be observed in font, color, and size. However, when a word is read, its meaning is independent of the style in which it has been written or rendered. In this paper, we investigate whether, and how, the style in…

2. ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation

   Tengda Guo (Tsinghua University), Jiawang Hu (StepFun), Yifei Tao (Nanyang Technological University) et al. · arXiv:2610.22408

   AI summaryECHOES compresses visual history into recursive textual states using privileged multi-image teacher distillation, achieving long‑horizon GUI performance without storing past screenshots. Experiments on four benchmarks and ECHOESBMK show superior task success, memory fidelity, and efficient inference, with zero‑shot transfer.

3. CF-NODE: Counterfactual Spatial Node Intervention for Test-Time Adaptation Under Domain Shift, Evaluated Across Five Diabetic Retinopathy Datasets

   Sabarish G, Deepak Parashar · arXiv:2610.22437

   AI summaryCF‑NODE adapts a frozen DR grading model to each test image by counterfactually attenuating the most influential spatial nodes and calibrating the grade distribution based on the stability of the response. It achieves higher AUC and macro‑F1 than TENT, EATA, SAR, and T3A across five external datasets.

4. An Agentic ENF Framework for Forensic Video Authentication Against Generative AI

   Ryan Wyton (Southampton Solent University), Bacha Rehman (Southampton Solent University) · arXiv:2610.22440

   AI summaryThe paper proposes an ENF-based forensic framework that fuses audio and rolling‑shutter video signals to authenticate video timestamps. It achieves 93% audio and 100% video real‑recording detection, 1.0 AUC for claimed‑time verification, and cuts runtime by 54% with no accuracy loss.

5. Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning

   Yizhe Zhang, Xujia Wang, Songqi Tang et al. · arXiv:2610.22444

   AI summaryEvent-STVG introduces a new spatio‑temporal grounding task that requires long‑form event reasoning, role inference, and multi‑target localization across 4,016 annotated events. State‑of‑the‑art MLLMs achieve only 5.4% m‑vIoU, highlighting the challenge.

6. Erase the Mark, Preserve the Motion: Temporally Consistent Removal of Video Watermarks

   Yifan Liao, Zhen Sun, Zhicong Huang et al. · arXiv:2610.22457

   AI summaryPoseRegen removes video watermarks by applying a shared viewpoint transformation and reconstructing with a frozen DiT model conditioned on pose, preserving identity and motion. It achieves near‑random watermark accuracy (0.505 bits) and outperforms baselines by 9.34% on motion consistency metrics.

7. Has GPT-6-Astra Achieved Embodied Intelligence?

   SenseNova, Joint Research Team, : et al. · arXiv:2610.22471

   AI summaryGPT‑6‑Astra shows unprecedented performance on 12 embodied robotics benchmarks, achieving high success rates with in‑context learning and a System 2/1 control loop. However, it still struggles with costly demonstration collection, real‑time constraints, and long‑horizon robustness, limiting practical embodied intelligence.

8. A Unified Theory of Image Scaling: Point Spread Functions, Discrete Kernels, and Scale-Spaces

   Assaf Shocher · arXiv:2610.22472

   AI summaryThis paper derives closed‑form image‑scaling operators that exactly match optical zoom by using the camera’s point spread function as the base kernel. The operators compose multiplicatively, form a Moore‑Penrose pair, and enable faithful, invertible up/down scaling, improving stability of super‑resolution and classification.

9. Star Light, Star Bright: Can Multimodal Models Parse an Astronomy Visual Question Answering Dataset Right?

   J. P. Naiman · arXiv:2610.22474

   AI summaryThis paper introduces a new astronomy visual question answering dataset containing over two thousand figures and 47,000 questions to evaluate large multimodal models. Testing reveals that current low-cost proprietary models frequently fail at these domain-specific visual reasoning tasks.

10. JAVIS: Joint Audio-Visual Interaction and Streaming for Real-Time Conversational Agents

    Junseok Ahn, Jaehyun Park, Jeongsoo Choi et al. · arXiv:2610.22499

    AI summaryThe paper introduces JAVIS, an end-to-end streaming framework that uses a shared representation to generate synchronized speech and facial expressions for real-time conversational agents. JAVIS achieves significantly lower latency and superior lip synchronization compared to existing baselines.

11. KFS$^4$LAM: Adaptive KeyFrame Selection with Structure-Aware Mapping for Endoscopic Gaussian Splatting SLAM

    Xinwei Ju, Zhehua Mao, Tianyi Song et al. · arXiv:2610.22505

    AI summaryThe paper introduces KFS4LAM, an endoscopic Gaussian Splatting SLAM system that uses motion-guided adaptive keyframe selection and structure-aware mapping. It improves trajectory accuracy, depth estimation, and rendering quality on challenging colonoscopy datasets compared to existing baselines.

12. TriSAR-3D and PaintFusion: A Novel Trimodal LiDAR-RGB-Thermal Dataset and Method for 3D Person Detection in Search and Rescue

    Markus Kramer, Jan Wachenfeld, Oskar von Stryk · arXiv:2610.22509

    AI summaryThe authors introduce TriSAR-3D, a trimodal LiDAR-RGB-thermal dataset for search and rescue, and PaintFusion, a faster and more accurate 3D person detection method that outperforms existing baselines by effectively fusing multimodal data.

13. Towards Reconstructing Geographically Diverse Architecture with 3D Foundation Models

    Nicky Kriplani, Yiwen Zhang, Angelina Wang et al. · arXiv:2610.22514

    AI summaryThe authors introduce ArchWorld, a benchmark of geographically diverse architectural landmarks, to analyze performance disparities in 3D foundation models. They find that model errors stem from region-specific architectural features and propose an intervention scheme that improves overall accuracy while reducing geographic performance gaps.

14. Learning the Shared Latent Structure of Cryo-EM Inverse Problems

    Nithin Chalapathi, Theo Sternlieb, Ethan Han et al. · arXiv:2610.22523

    AI summaryThis paper introduces EMber, a family of vision transformers pre-trained on millions of cryo-EM images using complementary global and masked patch objectives. EMber improves conformational and compositional discrimination by over 14 percentage points and enables reference-quality reconstructions in fewer epochs.

15. FlashControl: One-Step Controllable Generator via Distillation-Friendly Single-Stream Teachers

    Ngan Nguyen, Dung Nguyen, Quan Dao et al. · arXiv:2610.22537

    AI summaryFlashControl finetunes selected layers of a standard text‑to‑image diffusion model to support spatial conditions, then distills this teacher into a one‑step model. The resulting FlashControl achieves single‑pass controllable generation with comparable quality to dual‑branch baselines.

16. CFC26: Building Evaluations for Deployment in Sonar-Based Fish Counting

    Madison Van Horn, Suzanne Stathatos, Sevan Brodjian et al. · arXiv:2610.22549

    AI summaryCFC26 expands sonar fish‑counting data to 11 rivers across 9 Pacific salmon systems and introduces realistic deployment‑style evaluation protocols. Two new metrics, nMNE and nMANE, uncover directional counting bias, revealing errors hidden by conventional frame‑based metrics.

17. Generative AI Produces and Transmits Beauty Bias

    AJ Alvero, Deepak Varuvel Dennison, Alexandra Crespin et al. · arXiv:2610.22553

    AI summaryAn audit of 27 text-to-image generators reveals that AI-generated faces are systematically more attractive than real people and reinforce racial and gender beauty biases. Furthermore, exposure to these attractive AI faces lowers the perceived hireability and trustworthiness of real job candidates.

18. Micro-Localization: High-Precision Mobile Robot Localization Using Microscopic Ground Texture

    Aaron Wilhelm, Marina de Albuquerque Marcon, Emith Uyanwatte et al. · arXiv:2610.22563

    AI summaryMicroscopic ground texture is used to achieve mobile robot localization with 28–73 µm RMS error, by linking magnification to precision and solving calibration, camera modeling, and focus issues. The system localizes all queries on various surfaces, matching photometric registration within ~9 µm.

19. Seeing Seizures: A Large-Scale Empirical Study of Video and EEG Deep Learning for Automated Rodent Seizure Detection and Severity Grading

    Alloy Das, Shreyan Ganguly, Nyzil Massey et al. · arXiv:2610.22566

    AI summaryThe study trains video and EEG models on 24,131 paired clips from 601 sessions of 20 rats to detect seizures and grade severity. R(2+1)D-18 and GRU achieve 0.968–0.974 F1 for detection; five‑class grading drops to 0.475–0.560, showing difficulty generalizing to new animals.

20. Alignment of visual brain encoding models with individual perceptual similarity judgments

    Daniel Chong, Haomiao Chen, Keith W Jamison et al. · arXiv:2610.22569

    AI summaryThis paper fine-tunes visual brain encoding models using human perceptual similarity judgments. The main result is that incorporating personalized similarity data produces models that best align with individual human perceptual judgments and capture intersubject correlations.

21. Stuck on Shuffle: On Attribute-Object Binding in Vision-Language Models

    Netanel Y. Tamir, Michael Green, Issar Tzachor et al. · arXiv:2610.22602

    AI summaryThe paper introduces SABER, a benchmark for attribute–object binding across all vision–language pairings, showing major failures in current models. It then fine‑tunes an MLLM embedder on synthetic attribute‑binding text, achieving state‑of‑the‑art SABER scores and improving other retrieval tasks.

22. Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition

    Jae Joong Lee, Bedrich Benes · arXiv:2610.22606

    AI summaryA new LiDAR benchmark, FOR-species20K, aggregates 25 datasets of 33 tree species. A ConvNeXtV2 model reaches 85% accuracy, but cross‑source generalization drops to 60%, showing the leaderboard score overestimates real-world transfer.

23. ProactGen: Learning Proactive Clarification for Video Generation

    Zhixiao Wang, Yang Chen, Yifan Xu et al. · arXiv:2610.22608

    AI summaryProactGen learns a questioning policy that proactively clarifies user intent during video generation, using reinforcement learning to ask the most valuable questions based on draft videos. The method boosts final alignment by at least 1.8× versus strongest baselines and outperforms a 4‑round baseline in 29% of its time.

24. Shift-Reliance Alignment: Ranking Vision-Language Models by the Cues They Trust

    Faisal Mohammad, Mohamed Younis · arXiv:2610.22610

    AI summaryShift-Reliance Alignment measures how much a VLM relies on cues like context or text by editing source images, then uses that reliance scaled by the shift (proportion of target questions misled) to rank models. SRA picks models that reduce error on misleading target sets, achieving up to 18.8% error versus 30.4% with source-only ranking.

25. LGGS: Longitudinal MRI Reconstruction via Guided Group Sparsity

    Nikola Janjušević, Hersh Chandarana, Yao Wang et al. · arXiv:2610.22643

    AI summaryThis paper introduces Longitudinal Guided Group Sparsity (LGGS), a deep-learning framework for accelerated MRI reconstruction that leverages prior patient scans. LGGS uses learned phase-invariant circulant attention to achieve state-of-the-art reconstruction performance while remaining robust to misregistration and pathological changes.

26. Is a Linear Probe Evidence of a Linear Representation?

    Nadir Kutluozen, Christian Internò, David Klindt · arXiv:2610.22655

    AI summaryLinear probes can be misleading: they succeed on any rich feature map, even random kernels, because they only capture a linear subspace. The paper introduces reverse predictivity, spectral concentration, and OOD tests to show pretrained transformers encode more complex, curved structures, not a purely linear representation.

27. Deep Learning for Deepfake Image Detection: Comparative Model Evaluation and Gradient-Based Visual Attribution

    Saad Korchi, Abderrahim Waga · arXiv:2610.22676

    AI summaryThis paper compares a custom convolutional neural network and a fine-tuned MobileNetV2 for deepfake image detection using a 140k face dataset. The pretrained MobileNetV2 achieves a superior validation accuracy of 95.90% compared to 84.15% for the model trained from scratch.

28. EgoP$^{2}$C: Internalizing Plan-Perceive-Compose for Spatial Reasoning in Egocentric Video

    Chaoyu Li, Arka Sadhu, Yuliang Li et al. · arXiv:2610.22681

    AI summaryEgoP2C internalizes a plan-perceive-compose reasoning process within a single vision-language model using multi-teacher on-policy distillation. This approach improves egocentric spatial reasoning accuracy across benchmarks by up to 9.2 percentage points while reducing inference FLOPs by over 65 percent.

29. Thinking from Above: Efficient Spatial Reasoning with Bird's-Eye-View Visual Tokens

    Faizan Siddiqui, Nirav Savaliya, Sahith Reddy Chada et al. · arXiv:2610.22700

    AI summaryCambrian‑BEV projects depth‑aware visual patches into a shared bird‑eye‑view grid, aggregating observations across time and viewpoints into 729 tokens. This spatial tokenization cuts LLM prefill time 2.9×, memory 3.6×, and boosts VSI‑Bench scores by 8.5 points.

30. HilbertGen-3D:Hilbert--Multifractal Conditioning for Topology-Aware 3D Generation

    Ziyi Zhang, Shixuan Li, Yorie Nakahira et al. · arXiv:2610.22720

    AI summaryHilbertGen‑3D introduces a Hilbert‑Multifractal Conditioning (HMC) scheme that encodes global 3D structure into a diffusion model, improving point‑cloud generation. On ShapeNet, it cuts 1‑NN error by up to 8.9 pp and boosts coverage by 5.9 pp versus TopoDiT‑3D.

31. Improving OCR Accuracy Through Multi-Classifier Verification of Visually Similar Characters

    Nuzhat Khan, Ab Al-Hadi Ab Rahman, Mohamad Anuar Kamaruddin et al. · arXiv:2610.22726

    AI summaryThe paper introduces a Character Verification Model that uses binary classifiers for visually confusable pairs to post‑process OCR outputs. It boosts accuracy by 3.95–14.81 pp across three engines, achieving 87.3 % correction of such errors with only 2.3 % false corrections.

32. OmniXRef: A Unified Framework for Multimodal Multi-Reference Video Editing

    Xiaoyan Cong, Tianyu Wang, Hyeonho Jeong et al. · arXiv:2610.22744

    AI summaryOmniXRef unifies text, image, and video references for multimodal video editing, using a single‑stream diffusion transformer and a reference‑bank interface. It outperforms task‑specific and commercial models on a new benchmark for camera, motion, content, style, and VFX transfer.

33. Cross-Person Identity Change Under Source-Coupled Multi-Human Generation: A Controlled UniPortrait Study

    Edrees Al-Edreesi, Ghaleb Al-Gaphari, Muneer Alsurori et al. · arXiv:2610.22746

    AI summaryThis paper investigates cross-person identity changes during multi-human personalization in UniPortrait. The authors find that native dual-ID generation better preserves identity for both protected and target individuals compared to single-target approaches, which cause broad facial perturbations.

34. MedSTP: Preserving Subspace and Transition Evidence for Medical VQA Token Pruning

    Mengchen Fan, Qizhen Lan, Xi Xiao et al. · arXiv:2610.22751

    AI summaryMedSTP introduces a training-free token pruning method for medical vision-language models that preserves spectral subspace evidence and uses a token-grid transition prior for boundary and texture sensitivity. It achieves strong performance on medical VQA benchmarks while reducing visual token sequences.

35. Input- and Feature-Space Correction for Cross-Hospital IHC Grading

    Jongwoon Lee, Jhair Gallardo, Christopher Kanan · arXiv:2610.22756

    AI summaryThis paper benchmarks input- and feature-space correction methods for cross-hospital immunohistochemistry grading using frozen pathology foundation models. The authors show that combining a proposed input transformation with feature-wise standardization significantly improves grading performance across hospitals.

36. Direct Conditional Parameterization for N-Dimensional Splatting

    Zhongpai Gao, Benjamin Planche, Van Nguyen Nguyen et al. · arXiv:2610.22778

    AI summaryThis paper introduces direct conditional parameterization for N-dimensional splatting, replacing costly on-the-fly covariance inversions with explicit per-effect parameters. The proposed method accelerates rendering up to 7.7x while matching or improving reconstruction quality across multiple benchmarks.

37. STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases

    Lieqi Liu, Rui gao, Jia-Chen Gu et al. · arXiv:2610.22795

    AI summaryThis paper introduces STRIDE, a benchmark for evaluating spatiotemporal reasoning in autonomous driving edge cases. The results show that current vision-language models struggle significantly with safety-critical trajectory prediction and scene-context awareness, achieving low accuracy despite strong general VQA capabilities.

38. Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition

    Se Un Park, Kyungkeon Chung, Huigeol Kim et al. · arXiv:2610.22811

    AI summaryChannel replay pruning removes 60–70% of YOLOv5/8 parameters without accuracy loss and outperforms other methods; a 5‑layer 18k‑parameter reader with 4‑bit weights/8‑bit activations matches FP32 accuracy, yielding 95.2% sign detection at 94 fps on an automotive SoC.

39. VISTA: A Multimodal Egocentric Benchmark for Goal-Oriented Perception and Assistance to Blind and Low-Vision Users

    Noé Mosca, Mohamed Rayan Barhdadi, Muhammad Hamza et al. · arXiv:2610.22822

    AI summaryVISTA provides a multimodal egocentric benchmark with 1,000 recordings from Project Aria glasses, offering six sensor streams and task annotations for blind/low‑vision assistance. Zero‑shot vision‑language models describe scenes well but fail to generate goal‑directed guidance, highlighting the need for better multimodal approaches.

40. VTrajEmbed: Aligning Text with Video Object Trajectories via Spatiotemporal Embeddings

    Ruizhi Chen, Zining Fan, Yudi Guo et al. · arXiv:2610.22842

    AI summaryVTrajEmbed decomposes videos into object trajectories and aligns each trajectory embedding with text. It outperforms Qwen3-VL by 15.5/18 % in trajectory‑level retrieval and reaches 49.48 % average R@1 across five benchmarks.

41. Variational and Decomposed method for Mitigating Bimodal Bias in Robust Visual Question Answering

    Yihan Zhao, Yuncheng Jiang · arXiv:2610.22845

    AI summaryThe paper analyzes how counting and non‑counting VQA questions are biased differently: counting bias comes from vision‑derived magnitudes, non‑counting bias mainly from language. It proposes VDB, a variational and decomposed bias mitigation method, which outperforms baselines on VQACPv1/2 and GQAOOD.

42. Measurement-Adaptive Light-Cone Transform for Robust Non-Line-of-Sight Imaging

    Lianfang Wang, Xueying Liu, Junnan Dong et al. · arXiv:2610.22847

    AI summaryThe paper introduces Measurement-Adaptive Light-Cone Transform (MALCT), a physics-guided framework for non-line-of-sight imaging that combines analytical and learned responses within a variational formulation. MALCT improves reconstruction quality and robustness under low-signal and limited-acquisition conditions while retaining the efficient Fourier-domain inversion of traditional methods.

43. HAPACT: A Benchmark For Human-Centric Physical Impact Localization in Movies

    Youngrae Kim, Yejin Jang, Eunho Kim et al. · arXiv:2610.22895

    AI summaryThe paper introduces HAPACT, a dataset of 2.8K movie clips with 20K fine‑grained human‑annotated impact frames, and defines a new task: temporal and SMPL‑X spatial localization of physical impacts on people. Results show existing models perform poorly, highlighting the task’s difficulty.

44. COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning

    Seungyoon Woo, Keonhee Park, Gunhee Kim · arXiv:2610.22907

    AI summaryCOMET trains a frozen VLM teacher to provide a shared image‑text feature reference, then uses tri‑affinity distillation to preserve multimodal geometry and a bandit‑guided routing of task experts to avoid interference and capacity limits. On X‑TAIL, this decoupled approach boosts continual VLM accuracy by up to 10 %.

45. MaskMerge: Structure-aware Token Merging for Efficient Instance Segmentation

    Bavesh Balaji, Letian Jiang, Alexander Wong et al. · arXiv:2610.22921

    AI summaryMaskMerge merges tokens using intermediate mask predictions to avoid fusing adjacent objects of the same class, improving mAP from 30.1 to 32.1 on PanNuke and 33.1 to 36.2 on iSAID while cutting GFLOPs by 33% and 9%, respectively.

46. ODSS-Net: Orthogonally Decoupled Shared-Specific Representations for Multi-Modal Object Re-Identification

    Zhifeng Ma, Yuhong Feng, Chongyi Huang et al. · arXiv:2610.22922

    AI summaryODSS‑Net learns orthogonally decoupled shared‑specific features for multi‑modal ReID: a shared structural module removes appearance bias, an orthogonal projection isolates identity cues, and a sample‑adaptive mixture fuses modalities; it outperforms baselines on three ReID benchmarks.

47. DICL-Qwen3VL: Decoupled Confidence Calibration for Interpretable Workwear Detection

    Meng Wang, Wu Zeng, Mei Li et al. · arXiv:2610.22929

    AI summaryDICL-Qwen3VL adapts Qwen3-VL with LoRA for workwear localization, decouples visual and reasoning confidence, and uses token entropy and perturbation sensitivity to calibrate. It achieves mAP@50 of 0.730 on ICSWD and 0.753 on WICWD, outperforming LoRA‑VLM baselines while providing explicit evidence and separate confidence scores.

48. MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models

    Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari et al. · arXiv:2610.22951

    AI summaryMambaAct replaces the transformer action expert in the flow‑matching VLA π₀.₅ with a compact Mamba mixer and a Context Compressor, reducing denoising latency by ~60% and improving LIBERO success by 1.08% while keeping a 10× smaller action expert.

49. Query-adaptive Recurrent Tokenization

    Bavesh Balaji, Pranav Gopinath, Pouya Sadeghi et al. · arXiv:2610.22957

    AI summaryQuery-Adaptive Recurrent Tokenization (QART) dynamically allocates visual tokens per query, steering visual representations toward query-relevant content and stopping when reconstruction error is low. It reduces token usage by 20‑25× while achieving 81.96% on Visual7W and 76.37% on VQAv2.

50. Stabilizing Illumination Conditioning in SSM-Based Low-Light Image Enhancement with Bottleneck Prompt Sidecars

    Wei Chen Ng, Qiyi Xie, Jian Ding Tan et al. · arXiv:2610.22958

    AI summaryThe paper introduces a bottleneck-side PromptAdapter that injects an estimated illumination map into an SSM-based low‑light enhancement model via prompt tokens, improving PSNR by up to 1.24 dB and boosting SSIM, LPIPS, and dark‑region metrics on LOL datasets.

51. ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark

    Joel Julin, Souraja Kundu, Liza Dahiya et al. · arXiv:2610.22964

    AI summaryANCHOR benchmarks multimodal reasoning by measuring whether explanations align with visual, temporal, and audio evidence. Across 12 MLLMs, models with similar answer accuracy differ widely in groundedness, revealing a significant alignment gap.

52. Sink-Sparse Hybrid Attention for Efficient Inference in T2I/T2V Diffusion Transformers

    Bingxin Xing, Yu Zhang, Jiatong Ding et al. · arXiv:2610.22975

    AI summaryThe paper replaces dense self‑attention in Diffusion Transformers with a train‑free Sink‑Sparse Hybrid Attention that reuses static sink masks and a lightweight pooled‑KV branch, achieving ~1.5× speedup while preserving image and video quality.

53. How Murky Before a VLM Knows It Cannot See? Accuracy, Confidence and Abstention on a Physical Visibility Axis

    Mochamad Rizal Fauzan, Irgi Surya · arXiv:2610.22981

    AI summaryThe study measures VLM accuracy, internal, verbalized confidence, and abstention on a physical visibility axis (fog and underwater) for 1000 COCO scenes. Accuracy drops to chance at 12–16 m fog, yet high confidence remains, revealing a blind‑spot width of at least 6× visibility.

54. UniField++: A Unified Agentic Framework with Loop Reasoning for MRI Field-Strength Enhancement

    Yiyang Lin, Chenhui Wang, Tianliang Yao et al. · arXiv:2610.22982

    AI summaryUniField++ introduces a multi-agent framework combining a vision-language planner and a flow-matching enhancer for MRI field-strength enhancement. Trained with radiologist annotations, it uses loop reasoning to target hallucination-prone regions, improving PSNR by 2.74 dB across transitions.

55. A100: A Hybrid Vision Foundation Model for Agricultural Plant Image Analysis

    Mayur Pratim Das, Ajai Dubey, Aksshay Mathew P et al. · arXiv:2610.22991

    AI summaryThe authors propose A100, a hybrid CNN-Transformer vision foundation model pre-trained via self-supervised learning on three million unlabeled agricultural images. A100 achieves superior performance across crop classification, disease detection, and phenotyping tasks while demonstrating high label efficiency.

56. BRICE: Brain-inspired Recurrent Illusory Contour Encoder for Point Cloud Completion

    Yunlei Wang, Dawei Li, Xue-song Tang · arXiv:2610.23001

    AI summaryThis paper introduces BRICE, a brain-inspired recurrent encoder that mimics visual illusory contours using predictive coding to complete missing 3D point clouds. Experiments show that integrating BRICE consistently enhances the performance of mainstream completion baselines across multiple datasets.

57. Semantic-Critical Regions for Video-to-Audio Generation

    Yun Fang, Wei Guo, Zhuonan Liang et al. · arXiv:2610.23021

    AI summarySCR-V2A localizes fine‑grained “semantic‑critical regions” in video that produce sound, encodes them with dual visual encoders, and conditions a lightweight audio generator on this representation. It outperforms prior V2A methods on distribution matching and audio quality while using few parameters.

58. Frame-Consistent Feed-Forward 3D Gaussian Splatting from Sparse Unposed Images

    Viruthshaan Vinasirajan, Ranga Rodrigo · arXiv:2610.23039

    AI summaryThe paper corrects two integration errors in a two‑stage feed‑forward Gaussian splatting pipeline (frozen VGGT for pose, TokenGS for reconstruction), enabling accurate pose handling without refinement. With the fixes, it achieves 22.47 PSNR, nearly COLMAP‑level, even from four sparse unposed images.

59. SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

    Kai Sun, Peibo Duan, Zhanglu Yan et al. · arXiv:2610.23046

    AI summarySpikeQ applies spike‑guided post‑training quantization for SNNs, selecting weight levels by minimizing spike errors and using a reset‑aware grid for membrane potentials to avoid threshold flips. It retains full‑precision accuracy within 2.59 points at 4 bits, shrinking memory 7.8× and energy 6.4×.

60. TriFold: Folding irregular 3D Gaussians into structured latents for 3D Scene Generation

    Yun Wang, Chiyu Wang, Fan Jiang et al. · arXiv:2610.23050

    AI summaryTriFold encodes irregular 3D Gaussian scenes into a fixed triplane latent, using cell‑wise learned queries and occupancy‑guided set prediction to reconstruct variable‑count primitives. It achieves high‑fidelity single‑image conditional generation of native 3D Gaussian scenes.

61. Domain-specific language alignment of multimodal embeddings towards explainable predictions

    Gabrielė Tijūnaitytė, Thijs L. van der Plas, Mohamed-Bachir Belaid et al. · arXiv:2610.23110

    AI summaryAligns any multimodal embedding to language using captioning of auxiliary data, a soft contrastive loss, and concept retrieval. Applied to Earth observation embeddings, it outperforms prior models on species‑distribution metrics and yields interpretable habitat associations.

62. Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation

    Xuanyu Liu, Hongyan Feng, Miao Pan et al. · arXiv:2610.23121

    AI summaryScene-Designer learns design intelligence for 3D indoor scene generation by alternating design and critique experiences using Scene-OTA-35K and Scene-OTC-15K, and introduces Critic‑guided GRPO to reduce trial‑and‑error, achieving state‑of‑the‑art physical validity, functional correctness, and layout coherence.

63. Reliability-Consistent Appearance Modeling for RGB-Infrared Multi-Object Tracking

    Bingyang Wang, Jiazuo Yu, Lu Zhang et al. · arXiv:2610.23125

    AI summaryRCA‑MOT keeps separate RGB, IR, and fused appearance states and uses a simple target‑scene estimator to compute branch‑wise reliabilities. These reliabilities regulate identity matching, boundary decisions, and memory updates, yielding +5.4 HOTA and +11.5 IDF1 gains on VT‑MOT and UniRTL.

64. From GAN-Based Models to GPT-5: A Taxonomy-Guided Study of Human Performance in Detecting Deepfake Face Images

    Zhiqing Guo, Lixin Jia, Luning Sun et al. · arXiv:2610.23144

    AI summaryThe paper introduces a taxonomy of visual artifacts for deepfake face images and evaluates human detection accuracy across GAN, diffusion, and GPT-5 generated fakes. It finds detection performance is model‑dependent, model‑specific training yields larger gains than general training, and accuracy remains stable over five months.

65. Knowledge-Guided Distributional Prompt Learning with Visual Hierarchy for Zero-Shot Anomaly Detection

    Jiyul Ham, Seongho Yoon, Yonggon Jung et al. · arXiv:2610.23172

    AI summaryHiDist models prompt tokens as Gaussian distributions, using normality references to generate anomaly prompts as deviations, and splits these into global and local representations for image‑level detection and pixel‑level localization, achieving strong cross‑dataset ZSAD performance.

66. MLLM Frame Subset Ensembling for Audio-Visual Video QA and MLLM-based Reranking for Ad-hoc Video Search in TRECVID 2025

    Andreas Goulas, Damianos Galanopoulos, Ioannis Patras et al. · arXiv:2610.23214

    AI summaryThe paper presents a two‑stage retrieval and reranking pipeline for TRECVID 2025: first stage uses VLMs with LLM‑generated query rephrasings, second stage applies an untrained MLLM to re‑score top shots, improving AVS rankings. For VQA, the team fine‑tunes an audio‑visual MLLM and uses frame‑subset ensembling to achieve the best AG performance and second‑best MC results.

67. RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars

    Haonan Han, Yukang Lin, Minliang Lin et al. · arXiv:2610.23230

    AI summaryRAP provides 25K verified paper‑poster pairs with HTML/CSS markup and a 150‑paper test set, enabling automatic render‑and‑compare evaluation. A retrieval‑based PosterRAG pipeline fills slots of retrieved templates, and an SFT+RL model can be trained on the same data.

68. X-Splat: Unified Gaussian Splatting for Joint RGB and X-ray 3D Reconstruction

    Wangze Xu, Chu Zhou, Mae Yamaguchi et al. · arXiv:2610.23237

    AI summaryX‑Splat jointly reconstructs RGB and X‑ray 3D scenes using a single set of Gaussian primitives, sharing geometry and only attaching appearance to surface‑proximal Gaussians. The unified model, supervised by both modalities and a depth consistency loss, improves alignment stability, RGB quality, and compactness over separate pipelines.

69. Kingfisher: Fast Personalized Visual Autoregressive Generation via Scale-Aware Prompt-Conditioned Scale Decomposition

    Dinh-Khoi Vo, Akihiro Sugimoto, Minh-Triet Tran et al. · arXiv:2610.23238

    AI summaryKingfisher personalizes a frozen visual autoregressive generator by decomposing early residual scale codes into content and style components conditioned on prompts, recombining them into a semantic prefix, and predicting remaining scales with self‑correction. It outperforms diffusion methods, saving 54× preparation time.

70. Can the Parts Fool the Test? Counterfactual Pairing Cycles for Relational OOD

    Yibo Zhou, Yirong Xiang · arXiv:2610.23248

    AI summaryThis paper introduces counterfactual pairing cycles, a blocked evaluation and training primitive that holds image and caption parts constant while swapping their pairings to eliminate superficial shortcuts. This approach successfully isolates relational out-of-distribution errors and improves vision-language model accuracy on complex pairings.

71. LonGaussian: Tracking Foundation Priors for Long-Context Feedforward Gaussian Reconstruction

    Zhoufeng Xie, Guanjun Wu, Shuai Zhang et al. · arXiv:2610.23250

    AI summaryLonGaussian is a streaming feed-forward framework that reconstructs 3D scenes from long video sequences by using multi-frame point tracks and depth consistency to merge redundant Gaussian predictions. It improves novel-view rendering quality on benchmark datasets while substantially reducing storage and rendering costs.

72. Expand-and-Prune: Test-Time Candidate-Space Adaptation for End-to-End Autonomous Driving

    Fabian Schmidt, Karol Fedurko, Markus Enzweiler et al. · arXiv:2610.23262

    AI summaryExpand‑and‑Prune expands a planner’s candidate set with intent‑guided and temporally propagated trajectories, then prunes them using prediction uncertainty at test time, achieving a 7.7 % lift in Driving Score on LangAuto and consistent gains across three other benchmarks.

73. SCALE: Improving Guidance-Free Visual Generation via Self-Contrastive Alignment for Likelihood Estimation

    Jiwan Hur, Dong-Jae Lee, Dongyeun Lee et al. · arXiv:2610.23263

    AI summaryThis paper introduces Self-Contrastive Alignment for Likelihood Estimation (SCALE), a guidance-free training framework that contrasts predictions on real and self-generated samples. SCALE achieves superior visual generation quality in a single forward pass, outperforming traditional classifier-free guidance methods without their computational overhead.

74. Agentic Image Generation in the Wild

    Yifan Li, Bowen Qin, Kun Zhou et al. · arXiv:2610.23269

    AI summaryThis paper introduces Savanna, a benchmark for generating images of recent, hard-to-find entities using web search, and proposes Ranger-8B, an agentic framework that outperforms frontier models by locating, verifying, and isolating visual references before generation.

75. Xiaomi-Lens-0 Technical Report

    Pei Fu, Yike Liu, Yue Sun et al. · arXiv:2610.23279

    AI summaryXiaomi-Lens-0 is a 4B multimodal model that predicts segmentation masks directly from its visual representations, enabling grounded chain-of-thought reasoning. It improves average referring comprehension by 5.6 points over Qwen3.5 while achieving state-of-the-art segmentation performance.

76. I2V-DETACH: Source Grounding Detachment for Unauthorized Image-to-Video Generation

    Chanhui Lee, Yeonghwan Song, Yewon Kang et al. · arXiv:2610.23287

    AI summaryThis paper introduces I2V-DETACH, an image immunization method that protects against unauthorized image-to-video generation by disrupting source-image cue propagation during denoising. The approach effectively prevents source-content preservation without degrading visual quality or temporal coherence, while demonstrating strong black-box transferability.

77. The Optimal Ranking Tradeoffs Between Two Ranking Scores in Two-Class Classification

    Sébastien Piérard, Adrien Deliège, Marc Van Droogenbroeck · arXiv:2610.23294

    AI summaryThe paper shows that many common classification metrics are special cases of a single parameterized ranking score. It derives an exact formula for the optimal tradeoff between any pair of such scores, minimizing Kendall‐distance variance over all possible rankings.

78. FlashReFocus: Interactive Image Refocusing in Seconds

    Ching-Heng Cheng, Chia-Ming Lee, Ming-Ching Chang et al. · arXiv:2610.23297

    AI summaryFlashReFocus uses a single diffusion step to recover detail and a lightweight renderer conditioned on a defocus map for interactive single-image refocusing. It achieves state-of-the-art visual quality while running hundreds of times faster than existing generative methods.

79. Look, Reason, Draw a Region: Recursive Refinement for Spatial Context Prediction

    Wajeeha Tahir, Anna Tegon, Lionnus Kesting et al. · arXiv:2610.23302

    AI summaryThe paper introduces Spatial-TRM, a compact, weight-tied recursive model that uses frozen vision-language features for spatial context prediction. It achieves a 29.51% context accuracy on the RoboSpatial-Home benchmark, significantly outperforming single-step models with the same parameter count.

80. CatchUp: Depth-Aware Re-Admission of Visual Tokens for Efficient Large Vision-Language Models

    Eunju Park, Uiwon Hwang · arXiv:2610.23324

    AI summaryCatchUp re‑admit visual tokens depth‑wise in LVLM decoders without training, using oracle states to approximate missed evolution. It keeps 88.9% fewer active tokens, preserving 92% of performance and cutting latency by 8% on six benchmarks.

81. TransNSNet: A Transformer-Enhanced Multi-Scale Adversarial Framework for CT-to-PET Image Synthesis

    Shaojun Li, Xiaoyu Deng, Ikuo Keshi · arXiv:2610.23325

    AI summaryTransNSNet uses a Transformer‑enhanced multi‑scale adversarial network to synthesize PET images from CT, achieving top performance on nine of ten metrics with a smaller parameter set. The improvements are modest and not statistically significant after correction.

82. STC-MGNet: Spatial-Temporal-Channel Attention-Enhanced Multimodal Graph Network for Pedestrian Intention Prediction

    Yuanzhe Li, Meng Wang · arXiv:2610.23356

    AI summarySTC‑MGNet predicts pedestrian crossing intent by fusing pose, visual, and ego‑vehicle motion data using spatial‑temporal‑channel attention and multi‑stage channel gating. It achieves 70%, 87%, and 88% accuracy on JAADbeh, JAADall, and PIE, outperforming baselines.

83. UniPolar: Joint Image and Polarimetric Restoration under Compound Degradations via Physical Consistency Learning

    Yang Jin, Guangyu Guo, Linghao Shen et al. · arXiv:2610.23360

    Polarimetric image restoration aims to recover polarization angle images and reliable polarimetric quantities from degraded observations. Beyond visual fidelity, it also requires preserving the physical relationships underlying reliable polarimetric estimation. However, existing…

84. MOSS-Image Technical Report

    MOSI. AI · arXiv:2610.23374

    AI summaryMOSS‑Image trains a 5B diffusion transformer with a new recipe to render long, bilingual text and precise layouts, and adds an agentic prompt‑enhancement module that translates user intent into detailed specs. It achieves 96.3% text coverage on LLTB and ranks second on RLA.

85. FloodVLM-Pix: Language-Guided Pixel-Level Flood Understanding

    Chengyang Zhang, Zhihao Wang, Zhiyu Long et al. · arXiv:2610.23377

    As climate change intensifies, floods are becoming more frequent and destructive, making accurate flood understanding an urgent priority. However, existing vision-language models perform poorly on flood-related tasks, suffering from attentional bias and inaccurate descriptions…

86. LithoWorld: An Endpoint World Model for Lithography Process-Window Recovery

    Botong Zhao, Fang Yu, Xinyuan Chen et al. · arXiv:2610.23408

    AI summaryLithoWorld is a vision-based world model that uses scanning electron microscopy images to predict process-parameter corrections for semiconductor lithography. It achieves an 88.5 percent success rate in recovering qualified pattern regions, outperforming existing baseline methods.

87. Can We Use Saliency Maps for the Evaluation of Generative Image Models?

    Eleonora Ristori, Luca Bindini, Marco Lippi et al. · arXiv:2610.23411

    AI summaryResearchers train a classifier per generative model to spot real versus fake images and use saliency maps to reveal model‑specific artifacts. The combined accuracy and saliency score aligns with human perception, eliminating the need for pretrained feature backbones.

88. GeoReason: Bridging Logical Reasoning and Spatial Fidelity in Remote Sensing Segmentation

    Cheng Fan, Han Wu, Tingzhang Luo et al. · arXiv:2610.23416

    AI summaryGeoReason injects the segmentation token [SEG] into the initial query to maintain spatial grounding and uses a location‑aware chain‑of‑thought with a patch‑to‑token alignment loss to keep the token aligned with remote‑sensing textures. The method outperforms current SOTA on Earthreason and RRSIS‑D.

89. mmPencil: Toward Writing-Style-Independent In-Air Handwriting Recognition via mmWave Radar and Large Vision-Language Model

    Yifan Guo, Zhu Wang, Qian Qin et al. · arXiv:2610.23419

    AI summarymmPencil maps mmWave radar trajectories to visual inputs for large vision-language models, enabling 3D in‑air handwriting recognition that is insensitive to writing style and environment. It achieves 97.6% accuracy on 200 words for seen users and 92.5% zero‑shot on 50 words for unseen users.

90. Momentum: Persistent Dynamics for Video World Models

    Nedko Savov, Stefan Maria Ailuro, Danda Pani Paudel et al. · arXiv:2610.23423

    Video world models must capture how a scene evolves beyond the current camera view to generate consistent observations when its content reappears. Existing approaches rely on observation history, static point clouds, or image caches constructed from past frames. These…

91. ShadowCast: On-the-Fly Context-Aware Shadow Augmentation for Remote Sensing Image Segmentation

    Yechan Kim, Dongho Yoon, Jongmin Joo et al. · arXiv:2610.23440

    AI summaryShadowCast synthesizes realistic shadows from segmentation labels and injects them during training, requiring no extra network or metadata. It improves semantic segmentation on Vaihingen, Potsdam, and LoveDA, outperforming baselines and boosting performance when combined with other augmentations.

92. TriINR: Unsupervised Atmospheric Turbulence Mitigation with Triple Implicit Neural Representations

    Yuka Ogino, Rikuto Otsuka, Yuho Shoji et al. · arXiv:2610.23454

    AI summaryTriINR represents a sharp image, tilt, and blur each as an implicit neural representation, allowing frame‑count‑independent turbulence modeling and memory‑efficient optimization. It achieves the best accuracy among unsupervised methods on synthetic and real long‑range datasets.

93. When Evidence Vanishes: Distilling Visual-Evidence Dependency in Medical Vision-Language Models

    Qizhen Lan, Yiling Ma, Ruiqi Chen et al. · arXiv:2610.23469

    Compact medical vision-language models (VLMs) can answer clinical questions correctly without reliably using the image, relying instead on question wording or learned priors. Standard knowledge distillation copies a teacher's answer preference on the original image, but it…

94. TAN: A Size- and Form-Adaptive Masked-Convolution Encoder for Tissue Graphs

    Jan Herschel, Anselm von Gladiss · arXiv:2610.23475

    Tissue graphs represent histology images as graphs over tissue-conforming superpixels, which keeps homogeneous tissue regions intact instead of cutting them into rectangular patches. To embed superpixels, current encoders crop them to bounding boxes and then pad, resize, or mask…

95. PONDER: Learning to Generate in the Right Order

    Julius Erbach, Dominik Narnhofer, Bernt Schiele et al. · arXiv:2610.23531

    We introduce PONDER, a method that learns in which order an image generator should fill in content and which of its earlier predictions to revise. Masked generative models construct a sample by repeatedly filling in subsets of missing variables. The order of these partial…

96. ReVision: Reprogramming Image Generators for Dense Visual Perception

    Yitong Yang, Yinglin Wang, Xuexin Liu et al. · arXiv:2610.23533

    AI summaryReVision uses a frozen multimodal encoder and a frozen image generator to produce instruction‑conditioned, low‑rank parameter updates per image–instruction pair, enabling segmentation, depth, and normal prediction without task heads. It matches state‑of‑the‑art dense‑perception results while avoiding cross‑task interference.

97. PreciseEdit: Strong Yet Stable Semantic Transport for One-Step Image Editing

    Jingyi Wang, Xiao Lu · arXiv:2610.23535

    Recent one-step text-to-image (T2I) models enable efficient image synthesis, opening up new possibilities for interactive Web-based image editing. However, existing one-step editing methods often construct a global transport field through simple vector operations between…

98. PanoBench: A Panoptic Occupancy Prediction Benchmark for Autonomous Driving

    Yi Feng, Zizhan Guo, Junwu E et al. · arXiv:2610.23537

    Panoptic occupancy unifies 3D geometry, semantics, and instance identities, but its development is limited by scarce instance-aware annotations. Existing benchmarks mainly provide semantic labels reconstructed from incomplete sensor observations, while panoptic datasets have…

99. Execution Before Memory: Diagnosing Memory Failures in Video World Models

    Weiyang Jin, Jiwen Yu, Sihan Xu et al. · arXiv:2610.23542

    A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not…

100. LiveSync: Iterative Self-Training for Lip-Sync Generation

     Huaishao Luo, Huan Zhang, Guanbo Bao et al. · arXiv:2610.23558

     AI summaryLiveSync trains a mask‑based lip‑sync model by iteratively generating high‑confidence pseudo pairs from real videos, mixing them with real data to refine the same network. The method yields top HDTF lip‑sync accuracy and competitive AIGC-LipSync results while keeping low inference latency.

101. From Adult Renal Masses to Pediatric Wilms Tumors: Evaluating Cross-Domain Generalization of Kidney Lesion Segmentation on MRI

     Sarah de Boer, Myrthe A. D. Buser, Hartmut Häntze et al. · arXiv:2610.23568

     Deep learning segmentation models for pediatric oncology are often trained on small, single-center cohorts, whereas much larger datasets exist for adult diseases, such as kidney cancer. Whether models trained on adult renal lesions generalize to pediatric Wilms tumors remains…

102. CLEAR: Cooperative Latent Flow for Visual Counterfactual Explanations with Accurate Region Preservation

     Yan Zeng, Changlu Guo, Oskar Kristoffersen et al. · arXiv:2610.23577

     Visual counterfactual explanations (VCEs) should alter a classifier's prediction while remaining realistic and preserving decision-irrelevant content. Image-space methods provide high-quality counterfactuals but are computationally costly, whereas latent-space methods are more…

103. Enhancing RAG-based Medical VQA with Gated Multimodal Representations

     Daniela Jimenez-Pano, Hugo Jair Escalante, Delia Irazú Hernández-Farías · arXiv:2610.23588

     AI summaryMultimodal RAG uses a gated unit to fuse image and question into a joint query, allowing evidence retrieval to exploit both modalities while keeping the Med-LVLM frozen. It outperforms single‑modality baselines and rivals state‑of‑the‑art MVQA results.

104. SemPurify: Semantics-Aware Data Purification for Black-box Backdoor Defense

     Jiawei Yu, Syed Irfan Ali Meerza, Yi Wu et al. · arXiv:2610.23589

     Backdoor attacks have emerged as a critical security threat in machine learning, where adversaries embed triggers within poisoned training samples to manipulate a model's predictions on targeted inputs while leaving its general behavior unaffected. This stealthy nature makes…

105. Learning to Discriminate Scene Structures Makes Self-Supervised Depth Learning Scalable

     Mengtan Zhang, Yuchuan Cui, Yu Ma et al. · arXiv:2610.23600

     Self-supervised monocular depth estimation is appealing for its potential to scale with unlabeled data, yet existing methods often degrade when trained on diverse scenes. This study finds that, under self-supervision, models tend to encode dataset-specific structural biases…

106. ADA 2025: A Multimodal Dataset for Water-Deficit Stress Detection in Sugar Beets

     Piper Powell, Robert Franz, Amos Smith et al. · arXiv:2610.23609

     Research and development into artificial intelligence (AI) and machine learning (ML) applications in crop management are hindered by a lack of suitable public datasets. Whereas applications in disease and pest detection benefit from a comparatively higher amount of available…

107. Not Every Mention Is a Claim: Rethinking Object Hallucination Detection in Vision-Language Models

     Hsin-Ling Hsu, Neeraja Kirtane, Kuan-Hao Huang · arXiv:2610.23610

     Object hallucination in vision-language models (VLMs) is usually measured by checking each object a caption names against the image. Most benchmarks and detectors alike treat every object mention as a claim that the object is present. This assumption fails when a caption names…

108. Stereo LWIR-RGB Automatic Labeling for Low-Light Camera-Based VRU Detection

     Dimitrios Bouzoulas, Pyry Jäppinen, Pejman Habibiroudkenar et al. · arXiv:2610.23621

     VRU detection in low-light conditions is a key component in ensuring VRU safety during the most collision prone time of the day. Currently, advances in low-light detection are limited due to the lack of large scale datasets and automated labeling methods. Most autolabeling…

109. Diversity Without Labels: Task-Agnostic Coreset Selection Across Perception and Action

     Julia Hindel, Benjamin Ramtoula, Kyuhwan Yeon et al. · arXiv:2610.23642

     While data collection for model training is relatively easy, curating high-quality datasets from large collections is inherently challenging. Coreset estimation methods address this problem by identifying the best data subsets to train models on. However, existing approaches…

110. A Language-Informed Loss for Visual Generation

     Suchir Agarwal, Keshigeyan Chandrasegaran, Ziyu Chen et al. · arXiv:2610.23653

     AI summaryThe paper introduces LILIE, an invertible image encoder trained with a vision‑language model to generate captions from noised features, enabling a language‑informed loss for text‑to‑image generation. Using this loss, JiT models achieve a 14‑point gFD reduction on GPIC.

111. FlexPart: Text-Controlled Semantics-Aware 3D Part Segmentation at Flexible Granularity

     Yiqun Zhao, Jiachen Xu, Yihao Zhi et al. · arXiv:2610.23659

     A 3D asset admits multiple valid part decompositions, while semantics-aware segmentation should reflect both the desired granularity and the parts a user wants to extract. Ideally, a practical segmentation model should strive for a simple principle: what you ask for is what you…

112. An Efficient Approach to Beetle Damage Detection of Banana Using Deep Learning

     Md. Mehedi Hasan Naeem, Ishtiyak Ahmmad Araf, Md. Arefin Haque Mahir et al. · arXiv:2610.23661

     Visible surface damage of banana fruits by Basilepta subcostata (the banana fruit scarring beetle) can reduce their market value. Here, manual inspection of Banana fruits is subjective and difficult to scale. This study presents a lightweight framework based on YOLOv8n deep…

113. PCDFusion: Proposal-Context-Detail Bayesian Rendering for Infrared-Visible Image Fusion

     Rui Liu, Lin Gu, Hesong Li et al. · arXiv:2610.23663

     Infrared-visible image fusion (IVIF) aims to combine visible and thermal infrared observations for human perception and downstream vision tasks. Visible and thermal infrared observations are physically complementary but unreliable in different ways. Existing deep IVIF methods…

114. Agentic Physics-Grounded Video Generation via Memory-Augmented In-Context Reasoning

     Xikang Guan, Shaocong Yang, Yanqi Bao et al. · arXiv:2610.23677

     Physical plausibility is essential for realistic video generation. Existing methods impose physical guidance before synthesis via simulations or refined prompts, but this can limit dynamic diversity, miss errors that emerge only after generation, and require repeated…

115. FedRAPP: Federated Test-Time Prompt Tuning for Vision-Language Models with Reliability-Aware Prompt Personalization

     Syed Md. Ahnaf Hasan, Md Akil Raihan Iftee, Sajib Mistry et al. · arXiv:2610.23705

     Federated prompt learning enables vision-language models like CLIP to collaboratively adapt lightweight soft prompts across distributed clients without sharing raw data. However, existing federated prompt learning methods are designed primarily for supervised settings with…

116. Synthetic Image Oversampling for Rare-Class Representation Learning in Deep Facial Phenotyping

     Miles Liao, Zhanliang Wang, Hongzhuo Chen et al. · arXiv:2610.23730

     Recent deep facial phenotyping methods can support the recognition of genetic syndromes, but patient images are often scarce and unevenly distributed across disorders. Synthetic facial images offer a way to supplement the training data without collecting additional patient…

117. ESC: Escaping 2D Diffusion Ambiguity with Geometric Constraints for Consistent Novel View Synthesis

     Ge Ya Luo, Daniel Severo, Brian Karrer et al. · arXiv:2610.23732

     Generating 3D-consistent novel views from 2D diffusion models remains a fundamental challenge in computer vision. In this work, we introduce a framework that enhances geometric alignment for monocular view synthesis under drastic camera changes by jointly generating a target…

118. Material and Defocus Cues for Occluded Object Localization in Single Hyperspectral Images: An Exploratory Study

     Ali Zia, Jun Zhou · arXiv:2610.23733

     Partial occlusion makes object detection difficult when visible regions have weak texture or similar appearance. Hyperspectral imaging provides material-dependent spectra and wavelength-dependent optical blur, which offer complementary evidence for separating objects. We…

119. Beyond Multi-Stage Pipelines: A Label-Correlation Ensemble for Multi-Label Retinal Fundus Disease Classification

     Md Asif Uddin, Jannatul Ferdous, Asif Islam et al. · arXiv:2610.23747

     Retinal fundus images frequently carry more than one disease, and public fundus datasets are heterogeneous and severely long-tailed. Clinical triage suggests a staged design, in which a screening gate, a disease-group router and group-specific classifiers operate in sequence,…

120. MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing

     Houyuan Chen, Fanzhou Wang, Xiangyu Fan et al. · arXiv:2610.23751

     AI summaryMTEdit constructs 500K multi‑turn image editing trajectories via packed trajectory canvas, reducing visual drift. Fine‑tuned Qwen‑Image‑MTEdit outperforms other open‑source models on MTEdit‑Bench, showing the dataset’s effectiveness for multi‑turn editing.

121. A Paired RGB-Sonar Dataset for Fish Monitoring with Consumer Sonar

     Mourad Derbedrosian, Michael Hobley, Pietro Perona · arXiv:2610.23759

     Imaging sonar enables underwater observation in conditions where optical imaging is ineffective due to darkness and turbid water. However, imaging-sonar systems previously used for ecological monitoring have typically been expensive, limiting their accessibility. Consumer…

122. Video Diffusion Avatars: Taming Video Diffusion Models into Controllable 3D Avatars

     Simon Giebenhain, Tobias Kirschstein, Liam Schoneveld et al. · arXiv:2610.23772

     AI summaryVideo Diffusion Avatars fine‑tune a video diffusion model on a single image to generate a photo‑realistic, multi‑view 3D avatar with controllable pose, expression, and camera viewpoint. The method beats state‑of‑the‑art single‑image 3D avatars, reducing face LPIPS by 29 % and improving perceptual quality by 66 %.

123. The Collection of a Real-World, Large-Scale Dataset for Automated Skill Evaluation in Central Line Dressing Change

     Ummey Tanin (Pam), Shannan K. Hamlin (Pam), Hsin-Mei Chen (Pam) et al. · arXiv:2610.23789

     Central line-associated bloodstream infections (CLABSIs) remain a major source of preventable morbidity and mortality in healthcare settings. Proper execution of the central line dressing change (CLDC) procedure is essential for infection prevention, yet current approaches to…

124. Persistent Planar Memory for Video World Models

     Yuze He, Bin Tan, Zelin Gao et al. · arXiv:2610.23795

     Video world models that generate and explore environments at scale need a persistent 3D memory of what they have produced. Plane primitives offer a natural candidate: a compact set of oriented planes captures the dominant structural elements (e.g., walls, floors, facades) of…

125. StopWatch: Goal Grounding for Object-goal Navigation

     Xinhao Liu, Yunsheng Tian, Yuheng Qiu et al. · arXiv:2610.23841

     Pretrained VLMs bring rich world knowledge and semantic priors to object-goal navigation, and fine-tuning a VLM into a navigation policy has become a common recipe. Despite this progress, success rates remain low, and it is unclear which capability these fine-tuned policies…

126. Rethinking the design space of Vanilla Pixel Diffusion Transformers

     Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava et al. · arXiv:2610.23854

     Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla diffusion transformer can generate directly in pixel space by replacing the…

127. Learning to Correct Geometry in Generated Videos

     Weijie Lyu, Tiancheng Shen, Jason Li et al. · arXiv:2610.23856

     Generated videos often lose geometric consistency under camera motion: structures warp under viewpoint change, parallax becomes inconsistent, and smooth camera movement may collapse into shot-like transitions. We argue that this failure is largely a data problem. In web-scale…