Computer Vision · Cross-listed from cs.LG
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
Monday 15 December
In modern manufacturing, Visual Anomaly Detection (VAD) is essential for automated inspection and consistent product quality. Yet, increasingly dynamic and flexible production environments introduce key challenges: First, frequent product changes in small-batch and on-demand manufacturing require rapid model updates. Second, legacy edge hardware lacks the resources to train and run large AI models. Finally, both anomalous and normal training data are often scarce, particularly for newly introduced product variations. We investigate on-device continual learning for unsupervised VAD with localization, extending the PatchCore to incorporate online learning for real-world industrial scenarios. The proposed method leverages a lightweight feature extractor and an incremental coreset update mechanism based on k-center selection, enabling rapid, memory-efficient adaptation from limited data while eliminating costly cloud retraining. Evaluations on an industrial use case are conducted using a testbed designed to emulate flexible production with frequent variant changes in a controlled environment. Our method achieves a 12% AUROC improvement over the baseline, an 80% reduction in memory usage, and faster training compared to batch retraining. These results confirm that our method delivers accurate, resource-efficient, and adaptive VAD suitable for dynamic and smart manufacturing.
Computer Vision
Soul: Breathe Life into Digital Human for High-fidelity Long-term Multimodal Animation
We propose a multimodal-driven framework for high-fidelity long-term digital human animation termed $\textbf{Soul}$, which generates semantically coherent videos from a single-frame portrait image, text prompts, and audio, achieving precise lip synchronization, vivid facial expressions, and robust identity preservation. We construct Soul-1M, containing 1 million finely annotated samples with a precise automated annotation pipeline (covering portrait, upper-body, full-body, and multi-person scenes) to mitigate data scarcity, and we carefully curate Soul-Bench for comprehensive and fair evaluation of audio-/text-guided animation methods. The model is built on the Wan2.2-5B backbone, integrating audio-injection layers and multiple training strategies together with threshold-aware codebook replacement to ensure long-term generation consistency. Meanwhile, step/CFG distillation and a lightweight VAE are used to optimize inference efficiency, achieving an 11.4$\times$ speedup with negligible quality loss. Extensive experiments show that Soul significantly outperforms current leading open-source and commercial models on video quality, video-text alignment, identity preservation, and lip-synchronization accuracy, demonstrating broad applicability in real-world scenarios such as virtual anchors and film production. Project page at https://zhangzjn.github.io/projects/Soul/
Read the full abstract
Computer Vision
Transform Trained Transformer for Accelerating Native 4K Video Generation
Native 4K (2176$\times$3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotemporal resolution increases, making it difficult for models to strike a balance between efficiency and quality. This paper proposes a novel Transformer retrofit strategy termed T3 ($\textbf{T}$ransform $\textbf{T}$rained $\textbf{T}$ransformer) that, without altering the core architecture of full-attention pretrained models, significantly reduces compute requirements by optimizing their forward logic. Specifically, $\textbf{T3-Video}$ introduces a multi-scale weight-sharing window attention mechanism and, via hierarchical blocking together with an axis-preserving full-attention design, can effect an "attention pattern" transformation of a pretrained model using only modest compute and data. Results on $\textbf{4K-VBench}$ show that $\textbf{T3-Video}$ substantially outperforms existing approaches: while delivering performance improvements (+4.29$\uparrow$ VQA and +0.08$\uparrow$ VTC), it accelerates native 4K video generation by more than 10$\times$. Project page at https://zhangzjn.github.io/projects/T3-Video
Read the full abstract
Computer Vision
PoseAnything: Universal Pose-guided Video Generation with Part-aware Temporal Coherence
Pose-guided video generation refers to controlling the motion of subjects in generated video through a sequence of poses. It enables precise control over subject motion and has important applications in animation. However, current pose-guided video generation methods are limited to accepting only human poses as input, thus generalizing poorly to pose of other subjects. To address this issue, we propose PoseAnything, the first universal pose-guided video generation framework capable of handling both human and non-human characters, supporting arbitrary skeletal inputs. To enhance consistency preservation during motion, we introduce Part-aware Temporal Coherence Module, which divides the subject into different parts, establishes part correspondences, and computes cross-attention between corresponding parts across frames to achieve fine-grained part-level consistency. Additionally, we propose Subject and Camera Motion Decoupled CFG, a novel guidance strategy that, for the first time, enables independent camera movement control in pose-guided video generation, by separately injecting subject and camera motion control information into the positive and negative anchors of CFG. Furthermore, we present XPose, a high-quality public dataset containing 50,000 non-human pose-video pairs, along with an automated pipeline for annotation and filtering. Extensive experiments demonstrate that Pose-Anything significantly outperforms state-of-the-art methods in both effectiveness and generalization.
Read the full abstract
Computer Vision
Test-Time Modification: Inverse Domain Transformation for Robust Perception
Generative foundation models contain broad visual knowledge and can produce diverse image variations, making them particularly promising for advancing domain generalization tasks. They can be used for training data augmentation, but synthesizing comprehensive target-domain variations remains slow, expensive, and incomplete. We propose an alternative: using diffusion models at test time to map target images back to the source distribution where the downstream model was trained. This approach requires only a source domain description, preserves the task model, and eliminates large-scale synthetic data generation. We demonstrate consistent improvements across segmentation, detection, and classification tasks under challenging environmental shifts in real-to-real domain generalization scenarios with unknown target distributions. Our analysis spans multiple generative and downstream models, including an ensemble variant for enhanced robustness. The method improves BDD100K-Night-Det mAP@50 from 10.2 to 31.8, ImageNet-R top-1 from 36.1 to 60.8, and DarkZurich mIoU from 28.6 to 46.3.
Read the full abstract
Computer Vision
IMILIA: interpretable multiple instance learning for inflammation prediction in IBD from H&E whole slide images
As the therapeutic target for Inflammatory Bowel Disease (IBD) shifts toward histologic remission, the accurate assessment of microscopic inflammation has become increasingly central for evaluating disease activity and response to treatment. In this work, we introduce IMILIA (Interpretable Multiple Instance Learning for Inflammation Analysis), an end-to-end framework designed for the prediction of inflammation presence in IBD digitized slides stained with hematoxylin and eosin (H&E), followed by the automated computation of markers characterizing tissue regions driving the predictions. IMILIA is composed of an inflammation prediction module, consisting of a Multiple Instance Learning (MIL) model, and an interpretability module, divided in two blocks: HistoPLUS, for cell instance detection, segmentation and classification; and EpiSeg, for epithelium segmentation. IMILIA achieves a cross-validation ROC-AUC of 0.83 on the discovery cohort, and a ROC-AUC of 0.99 and 0.84 on two external validation cohorts. The interpretability module yields biologically consistent insights: tiles with higher predicted scores show increased densities of immune cells (lymphocytes, plasmocytes, neutrophils and eosinophils), whereas lower-scored tiles predominantly contain normal epithelial cells. Notably, these patterns were consistent across all datasets. Code and models to partially replicate the results on the public IBDColEpi dataset can be found at https://github.com/owkin/imilia.
Read the full abstract
Computer Vision · Cross-listed from eess.IV
Self-Supervised Ultrasound Representation Learning for Renal Anomaly Prediction in Prenatal Imaging
Prenatal ultrasound is the cornerstone for detecting congenital anomalies of the kidneys and urinary tract, but diagnosis is limited by operator dependence and suboptimal imaging conditions. We sought to assess the performance of a self-supervised ultrasound foundation model for automated fetal renal anomaly classification using a curated dataset of 969 two-dimensional ultrasound images. A pretrained Ultrasound Self-Supervised Foundation Model with Masked Autoencoding (USF-MAE) was fine-tuned for binary and multi-class classification of normal kidneys, urinary tract dilation, and multicystic dysplastic kidney. Models were compared with a DenseNet-169 convolutional baseline using cross-validation and an independent test set. USF-MAE consistently improved upon the baseline across all evaluation metrics in both binary and multi-class settings. USF-MAE achieved an improvement of about 1.87% (AUC) and 7.8% (F1-score) on the validation set, 2.32% (AUC) and 4.33% (F1-score) on the independent holdout test set. The largest gains were observed in the multi-class setting, where the improvement in AUC was 16.28% and 46.15% in F1-score. To facilitate model interpretability, Score-CAM visualizations were adapted for a transformer architecture and show that model predictions were informed by known, clinically relevant renal structures, including the renal pelvis in urinary tract dilation and cystic regions in multicystic dysplastic kidney. These results show that ultrasound-specific self-supervised learning can generate a useful representation as a foundation for downstream diagnostic tasks. The proposed framework offers a robust, interpretable approach to support the prenatal detection of renal anomalies and demonstrates the promise of foundation models in obstetric imaging.
Read the full abstract
Computer Vision
A Domain-Adapted Lightweight Ensemble for Resource-Efficient Few-Shot Plant Disease Classification
Accurate and timely identification of plant leaf diseases is essential for resilient and sustainable agriculture, yet most deep learning approaches rely on large annotated datasets and computationally intensive models that are unsuitable for data-scarce and resource-constrained environments. To address these challenges we present a few-shot learning approach within a lightweight yet efficient framework that combines domain-adapted MobileNetV2 and MobileNetV3 models as feature extractors, along with a feature fusion technique to generate robust feature representation. For the classification task, the fused features are passed through a Bi-LSTM classifier enhanced with attention mechanisms to capture sequential dependencies and focus on the most relevant features, thereby achieving optimal classification performance even in complex, real-world environments with noisy or cluttered backgrounds. The proposed framework was evaluated across multiple experimental setups, including both laboratory-controlled and field-captured datasets. On tomato leaf diseases from the PlantVillage dataset, it consistently improved performance across 1 to 15 shot scenarios, reaching 98.23+-0.33% at 15 shot, closely approaching the 99.98% SOTA benchmark achieved by a Transductive LSTM with attention, while remaining lightweight and mobile-friendly. Under real-world conditions using field images from the Dhan Shomadhan dataset, it maintained robust performance, reaching 69.28+-1.49% at 15-shot and demonstrating strong resilience to complex backgrounds. Notably, it also outperformed the previous SOTA accuracy of 96.0% on six diseases from PlantVillage, achieving 99.72% with only 15-shot learning. With a compact model size of approximately 40 MB and inference complexity of approximately 1.12 GFLOPs, this work establishes a scalable, mobile-ready foundation for precise plant disease diagnostics in data-scarce regions.
Read the full abstract
Computer Vision
MineTheGap: Automatic Mining of Biases in Text-to-Image Models
Text-to-Image (TTI) models generate images based on text prompts, which often leave certain aspects of the desired image ambiguous. When faced with these ambiguities, TTI models have been shown to exhibit biases in their interpretations. These biases can have societal impacts, e.g., when showing only a certain race for a stated occupation. They can also affect user experience when creating redundancy within a set of generated images instead of spanning diverse possibilities. Here, we introduce MineTheGap - a method for automatically mining prompts that cause a TTI model to generate biased outputs. Our method goes beyond merely detecting bias for a given prompt. Rather, it leverages a genetic algorithm to iteratively refine a pool of prompts, seeking for those that expose biases. This optimization process is driven by a novel bias score, which ranks biases according to their severity, as we validate on a dataset with known biases. For a given prompt, this score is obtained by comparing the distribution of generated images to the distribution of LLM-generated texts that constitute variations on the prompt. Code and examples are available on the project's webpage.
Read the full abstract
Computer Vision
RecTok: Reconstruction Distillation along Rectified Flow
Visual tokenizers play a crucial role in diffusion models. The dimensionality of latent space governs both reconstruction fidelity and the semantic expressiveness of the latent feature. However, a fundamental trade-off is inherent between dimensionality and generation quality, constraining existing methods to low-dimensional latent spaces. Although recent works have leveraged vision foundation models to enrich the semantics of visual tokenizers and accelerate convergence, high-dimensional tokenizers still underperform their low-dimensional counterparts. In this work, we propose RecTok, which overcomes the limitations of high-dimensional visual tokenizers through two key innovations: flow semantic distillation and reconstruction–alignment distillation. Our key insight is to make the forward flow in flow matching semantically rich, which serves as the training space of diffusion transformers, rather than focusing on the latent space as in previous works. Specifically, our method distills the semantic information in VFMs into the forward flow trajectories in flow matching. And we further enhance the semantics by introducing a masked feature reconstruction loss. Our RecTok achieves superior image reconstruction, generation quality, and discriminative performance. It achieves state-of-the-art results on the gFID-50K under both with and without classifier-free guidance settings, while maintaining a semantically rich latent space structure. Furthermore, as the latent dimensionality increases, we observe consistent improvements. Code and model are available at https://shi-qingyu.github.io/rectok.github.io.
Read the full abstract
Computer Vision
Learning to Generate Cross-Task Unexploitable Examples
Unexploitable example generation aims to transform personal images into their unexploitable (unlearnable) versions before they are uploaded online, thereby preventing unauthorized exploitation of online personal images. Recently, this task has garnered significant research attention due to its critical relevance to personal data privacy. Yet, despite recent progress, existing methods for this task can still suffer from limited practical applicability, as they can fail to generate examples that are broadly unexploitable across different real-world computer vision tasks. To deal with this problem, in this work, we propose a novel Meta Cross-Task Unexploitable Example Generation (MCT-UEG) framework. At the core of our framework, to optimize the unexploitable example generator for effectively producing broadly unexploitable examples, we design a flat-minima-oriented meta training and testing scheme. Extensive experiments show the efficacy of our framework.
Read the full abstract
Computer Vision
USTM: Unified Spatial and Temporal Modeling for Continuous Sign Language Recognition
Continuous sign language recognition (CSLR) requires precise spatio-temporal modeling to accurately recognize sequences of gestures in videos. Existing frameworks often rely on CNN-based spatial backbones combined with temporal convolution or recurrent modules. These techniques fail in capturing fine-grained hand and facial cues and modeling long-range temporal dependencies. To address these limitations, we propose the Unified Spatio-Temporal Modeling (USTM) framework, a spatio-temporal encoder that effectively models complex patterns using a combination of a Swin Transformer backbone enhanced with lightweight temporal adapter with positional embeddings (TAPE). Our framework captures fine-grained spatial features alongside short and long-term temporal context, enabling robust sign language recognition from RGB videos without relying on multi-stream inputs or auxiliary modalities. Extensive experiments on benchmarked datasets including PHOENIX14, PHOENIX14T, and CSL-Daily demonstrate that USTM achieves state-of-the-art performance against RGB-based as well as multi-modal CSLR approaches, while maintaining competitive performance against multi-stream approaches. These results highlight the strength and efficacy of the USTM framework for CSLR. The code is available at https://github.com/gufranSabri/USTM
Read the full abstract
Computer Vision
Computer vision training dataset generation for robotic environments using Gaussian splatting
This paper introduces a novel pipeline for generating large-scale, highly realistic, and automatically labeled datasets for computer vision tasks in robotic environments. Our approach addresses the critical challenges of the domain gap between synthetic and real-world imagery and the time-consuming bottleneck of manual annotation. We leverage 3D Gaussian Splatting (3DGS) to create photorealistic representations of the operational environment and objects. These assets are then used in a game engine where physics simulations create natural arrangements. A novel, two-pass rendering technique combines the realism of splats with a shadow map generated from proxy meshes. This map is then algorithmically composited with the image to add both physically plausible shadows and subtle highlights, significantly enhancing realism. Pixel-perfect segmentation masks are generated automatically and formatted for direct use with object detection models like YOLO. Our experiments show that a hybrid training strategy, combining a small set of real images with a large volume of our synthetic data, yields the best detection and segmentation performance, confirming this as an optimal strategy for efficiently achieving robust and accurate models.
Read the full abstract
Computer Vision
End2Reg: Learning Task-Specific Segmentation for Markerless Registration in Spine Surgery
Intraoperative navigation in spine surgery demands millimeter-level accuracy. Currently, this is achieved through radiation-intensive intraoperative imaging and bone-anchored markers that are invasive and disrupt surgical workflow. Markerless RGB-D registration methods offer a promising alternative. However, existing approaches rely on weak segmentation labels to isolate relevant anatomical structures, potentially propagating errors through the registration process. We present End2Reg, an end-to-end deep learning framework that jointly optimizes segmentation and registration, eliminating the need for segmentation labels and manual steps. The network learns task-specific segmentation masks optimized for registration, guided solely by the registration objective without explicit segmentation supervision. End2Reg achieves state-of-the-art performance on ex- and in-vivo benchmarks, reducing median Target Registration Error by 32% and mean Root Mean Square Error by 61%, while maintaining robust performance under partial occlusions. Ablation results confirm that end-to-end optimization significantly improves registration accuracy. Overall, End2Reg advances towards fully automatic, markerless intraoperative navigation. Code and interactive visualizations are available at: https://lorenzopettinari.github.io/end-2-reg/.
Read the full abstract
Computer Vision
rNCA: Self-Repairing Segmentation Masks
Accurately predicting topologically correct masks remains a difficult task for general segmentation models, which often produce fragmented or disconnected outputs. Fixing these artifacts typically requires hand-crafted refinement rules or architectures specialized to a particular task. Here, we show that Neural Cellular Automata (NCA) can be directly re-purposed as an effective refinement mechanism, using local, iterative updates guided by image context to repair segmentation masks. By training on imperfect masks and ground truths, the automaton learns the structural properties of the target shape while relying solely on local information. When applied to coarse, globally predicted masks, the learned dynamics progressively reconnect broken regions, prune loose fragments and converge towards stable, topologically consistent results. We show how refinement NCA (rNCA) can be easily applied to repair common topological errors produced by different base segmentation models and tasks: for fragmented retinal vessels, it yields 2-3% gains in Dice/clDice and improves Betti errors, reducing $\beta_0$ errors by 60% and $\beta_1$ by 20%; for myocardium, it repairs 61.5% of broken cases in a zero-shot setting while lowering ASSD and HD by 19% and 16%, respectively. This showcases NCA as effective and broadly applicable refiners.
Read the full abstract
Computer Vision
A Study of the Design Space of Motion and Disocclusion Control in Video Generation
Disocclusion occurs when object movement reveals previously hidden content. Existing image-to-video methods provide different forms of motion and disocclusion control, yet the design space they span remains poorly understood. In this work, we systematically explore an algorithmic design space spanning motion specification (text-based versus tracking-based) and disocclusion specification (text-based versus image-based). To facilitate this study, we construct a benchmark consisting of both synthetic scenes and captured scenes, covering rigid and deformable motions, as well as static and dynamic disoccluded objects. We compare existing and adapted methods across this design space to analyze the strengths and limitations of different control strategies for handling disocclusions during motion. Our study reveals that tracking-based guidance improves motion accuracy, while image-based guidance improves the fidelity of newly revealed content. However, jointly controlling the appearance and dynamics of disoccluded objects remains a challenging open problem, particularly in scenarios where newly revealed objects undergo their own motion after becoming visible.
Read the full abstract
Computer Vision
Unlocking Generalization in Polyp Segmentation with DINO Self-Attention "keys"
Automatic polyp segmentation is crucial for improving the clinical identification of colorectal cancer (CRC). While Deep Learning (DL) techniques have been extensively researched for this problem, current methods frequently struggle with generalization, particularly in data-constrained or challenging settings. Moreover, many existing polyp segmentation methods rely on complex, task-specific architectures. To address these limitations, we present a framework that leverages the intrinsic robustness of DINO self-attention "key" features for robust segmentation. Unlike traditional methods that extract tokens from the deepest layers of the Vision Transformer (ViT), our approach leverages the key features of the self-attention module with a simple convolutional decoder to predict polyp masks, resulting in enhanced performance and better generalizability. We validate our approach using a multi-center dataset under two rigorous protocols: Domain Generalization (DG) and Extreme Single Domain Generalization (ESDG). Our results, supported by a comprehensive statistical analysis, demonstrate that this pipeline achieves state-of-the-art (SOTA) performance, significantly enhancing generalization, particularly in data-scarce and challenging scenarios. While avoiding a polyp-specific architecture, we surpass well-established models like nnU-Net and UM-Net. Additionally, we provide a systematic benchmark of the DINO framework's evolution, quantifying the specific impact of architectural advancements on downstream polyp segmentation performance.
Read the full abstract
Computer Vision
Automated User Identification from Facial Thermograms with Siamese Networks
The article analyzes the use of thermal imaging technologies for biometric identification based on facial thermograms. It presents a comparative analysis of infrared spectral ranges (NIR, SWIR, MWIR, and LWIR). The paper also defines key requirements for thermal cameras used in biometric systems, including sensor resolution, thermal sensitivity, and a frame rate of at least 30 Hz. Siamese neural networks are proposed as an effective approach for automating the identification process. In experiments conducted on a proprietary dataset, the proposed method achieved an accuracy of approximately 80%. The study also examines the potential of hybrid systems that combine visible and infrared spectra to overcome the limitations of individual modalities. The results indicate that thermal imaging is a promising technology for developing reliable security systems.
Read the full abstract
Computer Vision
Face Identity Unlearning for Retrieval via Embedding Dispersion
Face recognition systems rely on learning highly discriminative and compact identity clusters to enable accurate retrieval. However, as with other surveillance-oriented technologies, such systems raise serious privacy concerns due to their potential for unauthorized identity tracking. While several works have explored machine unlearning as a means of privacy protection, their applicability to face retrieval - especially for modern embedding-based recognition models - remains largely unexplored. In this work, we study the problem of face identity unlearning for retrieval systems and present its inherent challenges. The goal is to make selected identities unretrievable by dispersing their embeddings on the hypersphere and preventing the formation of compact identity clusters that enable re-identification in the gallery. The primary challenge is to achieve this forgetting effect while preserving the discriminative structure of the embedding space and the retrieval performance of the model for the remaining identities. To address this, we evaluate several existing approximate class unlearning methods (e.g., Random Labeling, Gradient Ascent, Boundary Unlearning, and other recent approaches) in the context of face retrieval and propose a simple yet effective dispersion-based unlearning approach. Extensive experiments on standard benchmarks (VGGFace2, CelebA) demonstrate that our method achieves superior forgetting behavior while preserving retrieval utility.
Read the full abstract
Computer Vision
KlingAvatar 2.0 Technical Report
Avatar video generation models have achieved remarkable progress in recent years. However, prior work exhibits limited efficiency in generating long-duration high-resolution videos, suffering from temporal drifting, quality degradation, and weak prompt following as video length increases. To address these challenges, we propose KlingAvatar 2.0, a spatio-temporal cascade framework that performs upscaling in both spatial resolution and temporal dimension. The framework first generates low-resolution blueprint video keyframes that capture global semantics and motion, and then refines them into high-resolution, temporally coherent sub-clips using a first-last frame strategy, while retaining smooth temporal transitions in long-form videos. To enhance cross-modal instruction fusion and alignment in extended videos, we introduce a Co-Reasoning Director composed of three modality-specific large language model (LLM) experts. These experts reason about modality priorities and infer underlying user intent, converting inputs into detailed storylines through multi-turn dialogue. A Negative Director further refines negative prompts to improve instruction alignment. Building on these components, we extend the framework to support ID-specific multi-character control. Extensive experiments demonstrate that our model effectively addresses the challenges of efficient, multimodally aligned long-form high-resolution video generation, delivering enhanced visual clarity, realistic lip-teeth rendering with accurate lip synchronization, strong identity preservation, and coherent multimodal instruction following.
Read the full abstract
Computer Vision
ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement
While existing generation and unified models excel at general image generation, they struggle with tasks requiring deep reasoning, planning, and precise data-to-visual mapping abilities beyond general scenarios. To push beyond the existing limitations, we introduce a new and challenging task: creative table visualization, requiring the model to generate an infographic that faithfully and aesthetically visualizes the data from a given table. To address this challenge, we propose ShowTable, a pipeline that synergizes MLLMs with diffusion models via a progressive self-correcting process. The MLLM acts as the central orchestrator for reasoning the visual plan and judging visual errors to provide refined instructions, the diffusion execute the commands from MLLM, achieving high-fidelity results. To support this task and our pipeline, we introduce three automated data construction pipelines for training different modules. Furthermore, we introduce TableVisBench, a new benchmark with 800 challenging instances across 5 evaluation dimensions, to assess performance on this task. Experiments demonstrate that our pipeline, instantiated with different models, significantly outperforms baselines, highlighting its effective multi-modal reasoning, generation, and error correction capabilities.
Read the full abstract
Computer Vision
LINA: Learning INterventions Adaptively for Physical Alignment and Generalization in Diffusion Models
Diffusion models (DMs) have achieved remarkable success in image and video generation. However, they still struggle with (1) physical alignment and (2) out-of-distribution (OOD) instruction following. We argue that these issues stem from the models' failure to learn causal directions and to disentangle causal factors for novel recombination. We introduce the Causal Scene Graph (CSG) and the Physical Alignment Probe (PAP) dataset to enable diagnostic interventions. This analysis yields three key insights. First, DMs struggle with multi-hop reasoning for elements not explicitly determined in the prompt. Second, the prompt embedding contains disentangled representations for texture and physics. Third, visual causal structure is disproportionately established during the initial, computationally limited denoising steps. Based on these findings, we introduce LINA (Learning INterventions Adaptively), a novel framework that learns to predict prompt-specific interventions, which employs (1) targeted guidance in the prompt and visual latent spaces, and (2) a reallocated, causality-aware denoising schedule. Our approach enforces both physical alignment and OOD instruction following in image and video DMs, achieving state-of-the-art performance on challenging causal generation tasks and the Winoground dataset. Our project page is at https://opencausalab.github.io/LINA.
Read the full abstract
Computer Vision
CausalCLIP: Causally-Informed Feature Disentanglement and Filtering for Generalizable Detection of Generated Images
The rapid advancement of generative models has increased the demand for generated image detectors capable of generalizing across diverse and evolving generation techniques. However, existing methods, including those leveraging pre-trained vision-language models, often produce highly entangled representations, mixing task-relevant forensic cues (causal features) with spurious or irrelevant patterns (non-causal features), thus limiting generalization. To address this issue, we propose CausalCLIP, a framework that explicitly disentangles causal from non-causal features and employs targeted filtering guided by causal inference principles to retain only the most transferable and discriminative forensic cues. By modeling the generation process with a structural causal model and enforcing statistical independence through Gumbel-Softmax-based feature masking and Hilbert-Schmidt Independence Criterion (HSIC) constraints, CausalCLIP isolates stable causal features robust to distribution shifts. When tested on unseen generative models from different series, CausalCLIP demonstrates strong generalization ability, achieving improvements of 6.83% in accuracy and 4.06% in average precision over state-of-the-art methods.
Read the full abstract
Computer Vision
VideoASMR-Bench: Can AI-Generated ASMR Videos Fool VLMs and Humans?
With AI-generated videos increasingly indistinguishable from reality, current benchmarks primarily focus on broad semantic alignment and basic physical consistency, offering limited discriminative power for evaluating them. To address this, we introduce VideoASMR-Bench, a benchmark based on Autonomous Sensory Meridian Response (ASMR) videos that emphasizes fine-grained audio-visual perception and sensory immersion. This benchmark aims to answer two key questions: (i) Are today's video understanding models (VLMs) sensitive enough to detect AI-generated ASMR videos by recognizing minor visual, physical, or auditory artifacts? (ii) Can today's video generation models (VGMs) produce convincing ASMR videos with immersive experiences? This benchmark comprises a diverse set of 1,500 high-quality real ASMR videos curated from social media, alongside 2,235 synthetic counterparts generated by nine VGMs. Additionally, we open-source an extensible suite of prompts and reference images, enabling the benchmark to scale dynamically with future video models. Moreover, we design an automatic understanding-generation evaluation framework between VGMs and VLMs, where VGMs aim to produce realistic fake videos to fool the VLMs, while the VLMs seek to detect them, forming an adversarial game between the two parties. Our evaluation on VideoASMR-Bench reveals that even state-of-the-art VLMs, such as Gemini-3-Pro, fail to reliably detect AI-generated ASMR videos. Meanwhile, current frontier video generation models can produce ASMR videos that are difficult for VLMs to distinguish from real ones, while humans can still identify them relatively easily.
Read the full abstract
Computer Vision
CogniEdit: Dense Gradient Flow Optimization for Fine-Grained Image Editing
Instruction-based image editing with diffusion models has achieved impressive results, yet existing methods struggle with fine-grained instructions specifying precise attributes such as colors, positions, and quantities. While recent approaches employ Group Relative Policy Optimization (GRPO) for alignment, they optimize only at individual sampling steps, providing sparse feedback that limits trajectory-level control. We propose a unified framework CogniEdit, combining multi-modal reasoning with dense reward optimization that propagates gradients across consecutive denoising steps, enabling trajectory-level gradient flow through the sampling process. Our method comprises three components: (1) Multi-modal Large Language Models for decomposing complex instructions into actionable directives, (2) Dynamic Token Focus Relocation that adaptively emphasizes fine-grained attributes, and (3) Dense GRPO-based optimization that propagates gradients across consecutive steps for trajectory-level supervision. Extensive experiments on benchmark datasets demonstrate that our CogniEdit achieves state-of-the-art performance in balancing fine-grained instruction following with visual quality and editability preservation
Read the full abstract
Computer Vision · Cross-listed from cs.RO
Post-Training and Test-Time Scaling of Generative Agent Behavior Models for Interactive Autonomous Driving
Learning interactive motion behaviors among multiple agents is a core challenge in autonomous driving. While imitation learning models generate realistic trajectories, they often inherit biases from datasets dominated by safe demonstrations, limiting robustness in safety-critical cases. Moreover, most studies rely on open-loop evaluation, overlooking compounding errors in closed-loop execution. We address these limitations with two complementary strategies. First, we propose Group Relative Behavior Optimization (GRBO), a reinforcement learning post-training method that fine-tunes pretrained behavior models via group relative advantage maximization with human regularization. Using only 10% of the training dataset, GRBO improves safety performance by over 40% while preserving behavioral realism. Second, we introduce Warm-K, a warm-started Top-K sampling strategy that balances consistency and diversity in motion selection. Our Warm-K method-based test-time scaling enhances behavioral consistency and reactivity at test time without retraining, mitigating covariate shift and reducing performance discrepancies. Demo videos are available in the supplementary material.
Read the full abstract
Computer Vision
Toward Ambulatory Vision: Learning Visually-Grounded Active View Selection
Vision Language Models (VLMs) excel at visual question answering (VQA) but remain limited to snapshot vision, reasoning from static images. In contrast, embodied agents require ambulatory vision, actively moving to obtain more informative views. We introduce Visually Grounded Active View Selection (VG-AVS), a task that selects the most informative next viewpoint using only the visual information in the current image, without relying on scene memory or external knowledge. To support this task, we construct a synthetic dataset with automatically generated paired query-target views and question-answer prompts. We also propose a framework that fine-tunes pretrained VLMs through supervised fine-tuning (SFT) followed by RL-based policy optimization. Our approach achieves strong question answering performance based on viewpoint selection and generalizes robustly to unseen synthetic and real scenes. Furthermore, incorporating our learned VG-AVS framework into existing scene-exploration-based EQA systems improves downstream question-answering accuracy.
Read the full abstract
Computer Vision
STARCaster: Spatio-Temporal AutoRegressive Video Diffusion for Identity- and View-Aware Talking Portraits
This paper presents STARCaster, an identity-aware spatio-temporal video diffusion model that addresses both speech-driven portrait animation and dynamic viewpoint control, given an identity embedding or reference image, within a unified framework. Existing 2D speech-to-video diffusion models depend heavily on reference guidance, leading to limited motion diversity. At the same time, 3D-aware animation typically relies on inversion through pretrained tri-plane generators, which often leads to imperfect reconstructions and identity drift. We rethink reference- and geometry-based paradigms in two ways. First, we deviate from strict reference conditioning at pretraining by introducing softer identity constraints. Second, we address 3D awareness implicitly within the 2D video domain by leveraging the inherent multi-view nature of video data. STARCaster adopts a compositional approach progressing from ID-aware motion modeling, to audio-visual synchronization via lip reading-based supervision, and finally to novel view animation through temporal-to-spatial adaptation. To overcome the scarcity of 4D audio-visual data, we propose a decoupled learning approach in which view consistency and temporal coherence are trained independently. Comprehensive evaluations demonstrate that STARCaster generalizes effectively across tasks and identities, consistently surpassing prior approaches in different benchmarks.
Read the full abstract
Computer Vision
Ego-EXTRA: video-language Egocentric Dataset for EXpert-TRAinee assistance
We present Ego-EXTRA, a video-language Egocentric Dataset for EXpert-TRAinee assistance. Ego-EXTRA features 50 hours of unscripted egocentric videos of subjects performing procedural activities (the trainees) while guided by real-world experts who provide guidance and answer specific questions using natural language. Following a “Wizard of OZ” data collection paradigm, the expert enacts a wearable intelligent assistant, looking at the activities performed by the trainee exclusively from their egocentric point of view, answering questions when asked by the trainee, or proactively interacting with suggestions during the procedures. This unique data collection protocol enables Ego-EXTRA to capture a high-quality dialogue in which expert-level feedback is provided to the trainee. Two-way dialogues between experts and trainees are recorded, transcribed, and used to create a novel benchmark comprising more than 15k high-quality Visual Question Answer sets, which we use to evaluate Multimodal Large Language Models. The results show that Ego-EXTRA is challenging and highlight the limitations of current models when used to provide expert-level assistance to the user. The Ego-EXTRA dataset is publicly available to support the benchmark of egocentric video-language assistants: https://fpv-iplab.github.io/Ego-EXTRA/.
Read the full abstract
Computer Vision
POLAR: A Portrait OLAT Dataset and Generative Framework for Illumination-Aware Face Modeling
Face relighting aims to synthesize realistic portraits under novel illumination while preserving identity and geometry. However, progress remains constrained by the limited availability of large-scale, physically consistent illumination data. To address this, we introduce POLAR, a large-scale and physically calibrated One-Light-at-a-Time (OLAT) dataset containing over 200 subjects captured under 156 lighting directions, multiple views, and diverse expressions. Building upon POLAR, we develop a flow-based generative model POLARNet that predicts per-light OLAT responses from a single portrait, capturing fine-grained and direction-aware illumination effects while preserving facial identity. Unlike diffusion or background-conditioned methods that rely on statistical or contextual cues, our formulation models illumination as a continuous, physically interpretable transformation between lighting states, enabling scalable and controllable relighting. Together, POLAR and POLARNet form a unified illumination learning framework that links real data, generative synthesis, and physically grounded relighting, establishing a self-sustaining "chicken-and-egg" cycle for scalable and reproducible portrait illumination. Our project page: https://rex0191.github.io/POLAR/.