WACV 2026 论文列表

Understanding human affect via Valence-Arousal, Expressions, and Action Unit is essential for human-machine interaction. While recent multi-task learning (MTL) methods seek to unify these tasks, they overlook three key challenges: (i) the absence of unified modeling all three affective task types: regression, detection, and classification; (ii) reliance on complete annotations for all tasks, leaving disjoint single-task datasets underutilized; and (iii) task conflicts caused by Noisy Gradients, Negative Transfer (NT), and Task-specific Performance Misalignment (TPM). We introduce COIN, a novel two-stage MTL framework that bridges Cognitive Priors and Instance Semantics for robust training. First, we design a cognitively guided cross-task label induction strategy to propagate supervision under sparse annotations and mitigate NT, yielding strong task-specific CogXperts. Second, we introduce two complementary branches to address TPM: (i) Task-Specific Branch: transferring cognitive knowledge from task-optimal CogXperts to jointly optimize objectives under partial supervision, and (ii) Semantic Alignment Branch: enhancing instance-level semantic representations via Class-Conditioned and Instance-Adaptive Prompts. Experiments across six diverse datasets demonstrate COIN's robustness and generalization.

Active Speaker Detection (ASD) aims to identify who is currently speaking in each frame of a video. Most state-of-the-art approaches rely on late fusion to combine visual and audio features, but late fusion often fails to capture fine-grained cross-modal interactions, which can be critical for robust performance in unconstrained scenarios. In this paper, we introduce GateFusion, a novel architecture that combines strong pretrained unimodal encoders with a Hierarchical Gated Fusion Decoder (HiGate). HiGate enables progressive, multi-depth fusion by adaptively injecting contextual features from one modality into the other at multiple layers of the Transformer backbone, guided by learnable, bimodally-conditioned gates. To further strengthen multimodal learning, we propose two auxiliary objectives: Masked Alignment Loss (MAL) to align unimodal outputs with multimodal predictions, and Over-Positive Penalty (OPP) to suppress spurious video-only activations. GateFusion establishes new state-of-the-art results on several challenging ASD benchmarks, achieving 77.8% mAP (+9.4%), 86.1% mAP (+2.9%), and 96.1% mAP (+0.5%) on Ego4D-ASD, UniTalk, and WASD benchmarks, respectively, and delivering competitive performance on AVA-ActiveSpeaker. Out-of-domain experiments demonstrate the generalization of our model, while comprehensive ablations show the complementary benefits of each component.

Scalable Vector Graphics (SVGs) function both as visual images and as structured code that encode rich geometric and layout information, yet most methods rasterize them and discard this symbolic organization. At the same time, recent sentence embedding methods produce strong text representations but do not naturally extend to visual or structured modalities. We propose a training-free, instruction-guided multimodal embedding framework that uses a Multimodal Large Language Model (MLLM) to map text, raster images, and SVG code into an aligned embedding space. We control the direction of embeddings through modality-specific instructions and structural SVG cues, eliminating the need for learned projection heads or contrastive training. Our method has two key components: (1) Multimodal Explicit One-word Limitation (mEOL), which instructs the MLLM to summarize any multimodal input into a single token whose hidden state serves as a compact semantic embedding. (2) A semantic SVG rewriting module that assigns meaningful identifiers and simplifies nested SVG elements through visual reasoning over the rendered image, exposing geometric and relational cues hidden in raw code. Using a repurposed VGBench, we build the first text-to-SVG retrieval benchmark and show that our training-free embeddings outperform encoder-based and training-based multimodal baselines. These results highlight prompt-level control as an effective alternative to parameter-level training for structure-aware multimodal retrieval. Project page: https://scene-the-ella.github.io/meol/

The widespread use of synthetic image generation models and the challenges associated with authenticity preservation have fueled the demand for robust watermarking methods to safeguard authenticity and protect the copyright of synthetic images. Existing watermarking methods embed. Invisible signatures in synthetic images often compromise image quality and remain susceptible to multiple watermark removal attacks, including reconstruction and forgery methods. To overcome this issue, we propose a novel watermarking approach, AEON, which seamlessly integrates the watermark into the latent diffusion process and ensures it aligns with scene semantics in the final image. Unlike existing invisible in-diffusion watermarking and traditional hash-based methods, our approach adapts the neural synthesized hash-based watermark to the semantics of the generated image during the intermediate diffusion process instead of embedding traditional hashes with the initial noise. Our proposed approach a) modulates the noise sampling in each diffusion denoising iteration through a learnable watermark embedding, b) optimizes consistency, reconstruction, and similarity loss, enforcing local and global alignment between the watermark structure and the underlying image content, and c) generates a strong watermark by allowing late embedding of the watermark in the diffusion process. Empirical results demonstrate the effectiveness of the proposed approach in retaining quality and its robustness against cumulative adversarial attacks.

The task of style transfer for 3D Gaussian splats has been explored in many previous works, but these require reconstructing or fine-tuning the splat while incorporating style information or optimizing a feature extraction network on the splat representation. We propose a reconstruction- and optimization-free approach to stylizing 3D Gaussian splats, allowing for direct stylization on a .ply or .splat file without requiring the original camera views. This is done by generating a graph structure across the implicit surface of the splat representation. A feed-forward, surface-based stylization method is then used and interpolated back to the individual splats in the scene. This also allows for fast stylization of splats with no additional training, achieving speeds under 2 minutes even on CPU-based consumer hardware. We demonstrate the quality results this approach achieves and compare to other 3D Gaussian splat style transfer methods.

Evaluating text-to-image and text-to-video models is challenging due to a fundamental disconnect: established metrics fail to jointly measure visual quality and semantic alignment with text, leading to a poor correlation with human judgments. To address this critical issue, we propose cFreD, a general metric based on a Conditional Frechet Distance that unifies the assessment of visual fidelity and text-prompt consistency into a single score. Existing metrics such asFrechet Inception Distance (FID) capture image quality but ignore text conditioning while alignment scores such as CLIPScore are insensitive to visual quality. Furthermore, learned preference models require constant retraining and are unlikely to generalize to novel architectures or out-of-distribution prompts. Through extensive experiments across multiple recently proposed text-to-image models and diverse prompt datasets, cFreD exhibits a higher correlation with human judgments compared to statistical metrics, including metrics trained with human preferences. Our findings validate cFreD as a robust, future-proof metric for the systematic evaluation of text conditioned models, standardizing benchmarking in this rapidly evolving field. We plan to release and include in the Appendix an evaluation toolkit and benchmark.

Accurate road damage detection is crucial for timely infrastructure maintenance and public safety, but existing vision-only datasets and models lack the rich contextual understanding that textual information can provide. To address this limitation, we introduce RoadBench, the first multimodal benchmark for comprehensive road damage understanding. This dataset pairs high-resolution images of road damages with detailed textual descriptions, providing a richer context for model training. We also present RoadCLIP, a novel vision-language model that builds upon CLIP by integrating domain-specific enhancements. It includes a disease-aware positional encoding that captures spatial patterns of road defects and a mechanism for injecting road-condition priors to refine the model's understanding of road damages. We further employ a GPT-driven data generation pipeline to expand the image-text pairs in RoadBench, greatly increasing data diversity without exhaustive manual annotation. Experiments demonstrate that RoadCLIP achieves state-of-the-art performance on road damage recognition tasks, significantly outperforming existing vision-only models by 19.2%. These results highlight the advantages of integrating visual and textual information for enhanced road condition analysis, setting new benchmarks for the field and paving the way for more effective infrastructure monitoring through multimodal learning.

Vision Language Models (VLMs) can be trained more efficiently if training sets can be reduced in size. Recent work has shown the benefits of masking text during VLM training using a variety of strategies (truncation, random masking, block masking and syntax masking) and has reported syntax masking as the top performer. In this paper, we analyze the impact of different text masking strategies on the word frequency distribution of the training data, and show that this impact is connected to model success. Motivated by this finding, we propose a new frequency-based text masking approach, Contrastive Language-Image Pre-training with Word Frequency Masking (CLIPF). Extensive experiments demonstrate the advantages of CLIPF over syntax masking and other existing approaches, particularly when the number of input tokens decreases. We show that not only CLIPF, but also other existing masking strategies, outperform syntax masking when enough epochs are used during training, a finding of practical importance for selecting a text masking method for VLM training. Our code is available online.

Coresets are compact, representative subsets of large datasets. While coreset selection methods have been extensively investigated in image classification, their direct application to image quality assessment (IQA) is hindered by the incoherent and structurally distinct nature of content and quality representations in IQA tasks. To address this gap, we propose Q-Diverse coreset, a framework tailored for IQA. Our method begins by extracting dual-view embeddings that are both content-aware and quality-aware, capturing semantic and perceptual nuances. Rather than directly combining these heterogeneous features, we construct separate pairwise distance matrices and fuse them in the distance space. This fusion transforms into a graph-based structure from which spectral embeddings are derived. Finally, a geometric diversity-based sampling strategy is applied in the spectral space to select a coreset that maximizes representativeness. Notably, Q-Diverse operates in a label-free manner, making it especially valuable in IQA, where collecting quality score annotations is computationally expensive and time-consuming. Experimental results on seven IQA benchmarks demonstrate that Q-Diverse enables the effective training of deep learning-based IQA architectures, even with limited data, impressively retaining performance. It achieves SRCC and PLCC values within 0.045 and 0.042 of those obtained from full-data training, using only 10% of the dataset on average. Our results establish Q-Diverse as a coreset selection method that enables efficient dataset curation as well as training and fine-tuning deep learning-based IQA models. Code is available at: https://github.com/Arpita2012/Q-Diverse.

Despite the renewed interest in zero-shot visual classification due to the rise of Multimodal Large Language Models (MLLMs), the problem of evaluating free-form responses of autoregressive models remains a persistent challenge. Most existing works focus on language-only tasks or don't consider Multiple Choice Questions (MCQs) beyond 5-way options, both of which are critical capabilities to solve tasks in Fine-Grained Visual Classification (FGVC) where choice counts are in the hundreds to thousands and the choices are highly related. Furthermore, in this highly multi-way MCQ setting it is not clear how to extend LLM choice extraction to retrieval-based problems, where computing probabilities over the choice set is computationally costly. In this work we investigate nlg2choice, a simple two-stage method which first asks the MLLM an open-ended question for the task with minimal constraints, then uses text-only constrained decoding to predict the most likely choice. In retrieval settings, we compute the probability of the constrained response taking that choice with an early stopping method to significantly improve throughput. Our results show improvement over a suite of seven fine-grained visual datasets when evaluating in terms of classification and retrieval, and show that this performance holds over the various ways that users of LLMs can implement tasks in natural language.

We present FNOpt, a self-supervised cloth simulation framework that formulates time integration as an optimization problem and trains a resolution-agnostic neural optimizer parameterized by a Fourier neural operator (FNO). Prior neural simulators often rely on extensive ground-truth data or sacrifice fine-scale detail, and generalize poorly across resolutions and motion patterns. In contrast, FNOpt learns to simulate physically plausible cloth dynamics and achieves stable and accurate rollouts across diverse mesh resolutions and motion patterns without retraining. Trained only on a coarse grid with physics-based losses, FNOpt generalizes to finer resolutions, capturing fine-scale wrinkles and preserving rollout stability. Extensive evaluations on a benchmark cloth simulation dataset demonstrate that FNOpt outperforms prior learning-based approaches in out-of-distribution settings in both accuracy and robustness. These results position FNO-based meta-optimization as a compelling alternative to previous neural simulators for cloth; thus reducing the need for curated data and improving cross-resolution reliability.

Image geolocalization, in which an AI model traditionally predicts the precise GPS coordinates of an image, is a challenging task with many downstream applications. However, the user cannot utilize the model to further their knowledge beyond the GPS coordinates; the model lacks an understanding of the location and the conversational ability to communicate with the user. In recent days, with the tremendous progress of large multimodal models (LMMs), both proprietary and open-source researchers have attempted to geolocalize images using LMMs. However, the issues remain unaddressed; beyond general tasks, for more specialized downstream tasks, such as geolocalization, LMMs struggle. In this work, we propose solving this problem by introducing a conversational model, GAEA, that provides information regarding the location of an image as the user requires. No large-scale dataset exists that enables the training of such a model. Thus, we propose GAEA-1.4M, a comprehensive dataset comprising over 800k images and approximately 1.4M question-answer pairs, constructed by leveraging OpenStreetMap (OSM) attributes and geographical context clues. For quantitative evaluation, we propose a diverse benchmark, GAEA-Bench, comprising 3.5k image-text pairs to evaluate conversational capabilities equipped with diverse question types. We consider 11 state-of-the-art open-source and proprietary LMMs and demonstrate that GAEA significantly outperforms the best open-source model, LLaVA-OneVision, by 18.2% and the best proprietary model, GPT-4o, by 7.2%. Our dataset, model and codes are available https://ucf-crcv.github.io/GAEA.

In the realm of creative expression, not everyone possesses the gift of effortlessly translating their imaginative visions into flawless sketches. More often than not, the outcome resembles an abstract, perhaps even slightly distorted representation. The art of producing impeccable sketches is not only challenging but also a time-consuming process. Our work is the first of this kind in transforming abstract, sometimes deformed garment sketches into photorealistic catalog images, to empower the everyday individual to become their own fashion designer. We create Sketch2Stitch, a dataset featuring over 65,000 abstract sketch images generated from garments of Dress Code [??] and VITON HD [??], two benchmark datasets in the virtual try-on task. Sketch2Stitch is the first dataset in the literature to provide abstract sketches in the fashion domain. We propose a StyleGAN-based generative framework that bridges freehand sketching with photorealistic garment synthesis. We demonstrate that our framework allows users to sketch rough outlines and optionally provide color hints, producing realistic designs in seconds. Experimental results demonstrate, both quantitatively and qualitatively, that the proposed framework achieves superior performance against various baselines and existing methods on both subsets of our dataset. Our work highlights a pathway toward AI-assisted fashion design tools, democratizing garment ideation for students, independent designers, and casual creators.

Concept Bottleneck Models (CBMs) provide a basis for semantic abstractions within a neural network architecture. Such models have primarily been seen through the lens of interpretability so far, wherein they offer transparency by inferring predictions as a linear combination of semantic concepts. However, a linear combination is inherently limiting. So we propose the enhancement of concept-based learning models through propositional logic. We introduce a logic module that is carefully designed to connect the learned concepts from CBMs through differentiable logic operations, such that our proposed LogicCBM can go beyond simple weighted combinations of concepts to leverage various logical operations to yield the final predictions, while maintaining end-to-end learnability. Composing concepts using a set of logic operators enables the model to capture inter-concept relations, while simultaneously improving the expressivity of the model in terms of logic operations. Our empirical studies on well-known benchmarks and synthetic datasets demonstrate that these models have better accuracy, perform effective interventions and are highly interpretable.

In this paper, we present ZonUI-3B, a lightweight Vision-Language Model (VLM) that can be fully trained on a single consumer-grade GPU (RTX 4090) while delivering performance comparable to significantly larger models on GUI grounding tasks. The model incorporates several key innovations: (i) combine cross-platform, multi-resolution dataset of 24K examples from diverse sources including mobile, desktop, and web GUI screenshots to effectively address data scarcity in high-resolution desktop environments; (ii) a two-stage fine-tuning strategy, where initial cross-platform training establishes robust GUI understanding, followed by specialized fine-tuning on high-resolution data to significantly enhance model adaptability; and (iii) data curation and redundancy reduction strategies, demonstrating that randomly sampling a smaller subset with reduced redundancy achieves performance comparable to larger datasets, emphasizing data diversity over sheer volume. Empirical evaluation on standard GUI grounding benchmarks--including ScreenSpot, ScreenSpot-v2, and the challenging ScreenSpot-Pro--highlights ZonUI-3B's exceptional accuracy, achieving 84.9% on ScreenSpot and 86.4% on ScreenSpot-v2, surpassing prior models under 4B parameters. Ablation studies validate the critical role of balanced sampling and two-stage fine-tuning in enhancing robustness, particularly in high-resolution desktop scenarios.

Unsupervised anomaly segmentation in multi-sequence MRI is a promising way to scale lesion screening, but existing reconstruction-based methods face three persistent issues: they fail to generalize across modalities, they depend on hand-crafted masking or paired translations, and they often require separate models with high inference cost. In this work, we take a stepwise approach to address these limitations. In the first stage, we fully fine-tune a diffusion model on healthy brain MRI slices pooled across T1, T2, and FLAIR, which produces anatomically consistent reconstructions. To further improve, we introduce a lightweight second stage where modality-specific LoRA adapters are trained on top of the pretrained diffusion backbone. A simple router automatically selects the right adapter for each input, effectively turning the system into a modality-aware, agentic-like framework. To further stabilize reconstructions, we incorporate a learnable latent-frequency mask that suppresses non-informative spectral components and preserves structural detail. This design allows the model to emphasize healthy anatomy while efficiently capturing modality-dependent contrasts. This two-stage strategy boosts Dice to 88% on BraTS2021 (FLAIR), achieving state-of-the-art performance. Experiments on BraTS2021, ISLES, and ATLAS datasets confirm that the approach consistently improves Dice and SSIM across all modalities, outperforming diffusion, masking, and cycle-based baselines, and offering a practical balance between accuracy and efficiency for clinical MRI anomaly detection. Source code is available at https://github.com/wafaAlghallabi/MRI-Router.

We present AnyBald, a novel framework for realistic hair removal from portrait images captured under diverse in-the-wild conditions. One of the key challenges in this task is the lack of high-quality paired data, as existing datasets are often low-quality, with limited viewpoint variation and diversity, making it difficult to handle real-world cases. To address this, we construct a scalable data-augmentation pipeline that synthesizes high-quality hair and non-hair image pairs across diverse real-world scenarios, enabling effective generalization and scalable supervision. With this enriched dataset, we present a new hair-removal framework that reformulates pretrained latent diffusion inpainting using learnable text prompts, removing the need for explicit masking at inference. In doing so, our model achieves natural hair removal with semantic preservation via implicit localization. To further enhance spatial precision, we introduce a regularization loss that guides the model to attend specifically to hair regions. Extensive experiments demonstrate that AnyBald outperforms in removing hair while preserving identity and semantics across various in-the-wild domains. Our project page is here: https://vision3d-lab.github.io/anybald/.

AI video generation is undergoing a revolution, with quality and realism advancing rapidly. These advances have led to a passionate scientific debate: Do video models learn "world models" that discover laws of physics---or, alternatively, are they merely sophisticated pixel predictors that achieve visual realism without understanding the physical principles of reality? We address this question by developing Physics-IQ, a comprehensive benchmark dataset that can only be solved by acquiring a deep understanding of various physical principles, like fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. We find that across a range of current models (Sora, Runway, Pika, Lumiere, Stable Video Diffusion, and VideoPoet), physical understanding is severely limited, and unrelated to visual realism. At the same time, some test cases can already be successfully solved. This indicates that acquiring certain physical principles from observation alone may be possible, but significant challenges remain. While we expect rapid advances ahead, our work demonstrates that visual realism does not imply physical understanding. Our anonymous project page is [here](https://anonphysicsiq.github.io/); data and code are open-sourced; code is available from the supplementary materials.

Artistic text generation involves rendering textual content in visually creative and contextually appropriate designs, such as floral or geometric patterns. Despite advancements in Vision-Language Models (VLMs) like DALL*E and Qwen, maintaining both text accuracy and aesthetic alignment is a persistent challenge in this domain. This paper presents an analysis of the performance of several VLMs in generating artistic text, aiming to diagnose performance gaps and guide improvements in text accuracy and multimodal alignment. The objective of this study is to propose an analysis framework to evaluate and improve artistic text generation by assessing models across three key dimensions: character-level text accuracy, semantic similarity, and visual-context alignment. A custom dataset of 1,000 prompts, each specifying a word and artistic style to support systematic evaluation was created. We benchmarked the performance of DALL*E, Qwen-VL, and Qwen-2.5B on text accuracy by comparing the generated text to the original prompt and measuring perceptual alignment using CLIP embeddings. The results show that the DALL-E model outperforms the other models, in balancing text accuracy and style alignment. This study highlights the potential of adopting VLMs for artistic text generation. Furthermore, we present exploratory results using ControlNet, focusing on its potential to improve text accuracy in generated images. Our findings expose critical limitations in current VLMs and offer actionable insights for advancing multimodal generation architectures, datasets, and evaluation strategies, as well as provide a framework for improving text accuracy and aesthetic coherence in creative applications, paving the way for advancements in multimodal AI systems.