ICCV 2025 论文列表

Despite the success of adversarial training on small datasets, applying it to large-scale datasets like ImageNet remains challenging. Previous attempts using synthetic data show limited improvements. This work investigates the impact of synthetic data scaling, model scaling, and training strategies on adversarial training with ImageNet, providing deeper insights into large-scale robustness. During the process, we observe a notable phenomenon of loss oscillation, leading to adversarial overfitting, and propose strategies to mitigate it. Experimental results show that, under AutoAttack on ImageNet-1K, our method achieves a robust accuracy of 71.54%. Our findings highlight the crucial role of synthetic data and model scaling in enhancing adversarial robustness on large-scale benchmarks and provide a new direction for training robust visual representations at scale.

AI for tumor segmentation is limited by the lack of large, voxel-wise annotated datasets, which are hard to create and require medical experts. In our proprietary JHH dataset of 3,000 annotated pancreatic tumor scans, we found that AI performance stopped improving after 1,500 scans. With synthetic data, we reached the same performance using only 500 real scans. This finding suggests that synthetic data can steepen data scaling laws, enabling more efficient model training than real data alone. Motivated by these lessons, we created AbdomenAtlas 2.0---a dataset of 10,134 CT scans with a total of 13,223 tumor instances per-voxel manually annotated in six organs (pancreas, liver, kidney, colon, esophagus, and uterus) and 6,511 control scans. Annotated by 23 expert radiologists, it is several orders of magnitude larger than existing public tumor datasets. While we continue expanding the dataset, the current version of AbdomenAtlas 2.0 already provides a strong foundation---based on lessons from the JHH dataset---for training AI to segment tumors in six organs. It achieves notable improvements over public datasets, with a +7% DSC gain on in-distribution tests and +16% on out-of-distribution tests.

Large transformer-based models have made significant progress in generalizable novel view synthesis (NVS) from sparse input views, generating novel viewpoints without the need for test-time optimization. However, these models are constrained by the limited diversity of publicly available scene datasets, making most real-world (in-the-wild) scenes out-of-distribution. To overcome this, we incorporate synthetic training data generated from diffusion models, which improves generalization across unseen domains. While synthetic data offers scalability, we identify artifacts introduced during data generation as a key bottleneck affecting reconstruction quality. To address this, we propose a token disentanglement process within the transformer architecture, enhancing feature separation and ensuring more effective learning. This refinement not only improves reconstruction quality over standard transformers but also enables scalable training with synthetic data. As a result, our method outperforms existing models on both in-dataset and cross-dataset evaluations, achieving state-of-the-art results across multiple benchmarks while significantly reducing computational costs.

This work introduces Multimodal Context (MiCo), a scalable pretraining framework designed to advance omni-modal intelligence--an AI system capable of understanding and learning from multiple modalities to achieve universal representation learning. MiCo allows for efficient scaling of both the number of modalities and the volume of data, along with model parameters, during the pretraining phase. We evaluate the pretrained models across a diverse set of tasks, including: (i) single-modality perception benchmarks covering 10 distinct modalities, (ii) 25 cross-modal tasks spanning retrieval, question-answering, and captioning, and (iii) 18 large-scale multimodal language model benchmarks. MiCo consistently delivers state-of-the-art results, setting 37 new benchmarks across these tasks. The pretrained models, along with the collected datasets and codebase, will be made publicly available to support the development of omni-modal intelligence and broader research in multimodal learning.

Building general-purpose models that can effectively perceive the world through multimodal signals has been a long-standing goal. Current approaches involve integrating separately pre-trained components, such as connecting vision encoders to LLMs and continuing multimodal training. While such approaches exhibit remarkable sample efficiency, it remains an open question whether such late-fusion architectures are inherently superior. In this work, we revisit the architectural design of native multimodal models (NMMs)--those trained from the ground up on all modalities--and conduct an extensive scaling laws study, spanning 457 trained models with different architectures and training mixtures. Our investigation reveals no inherent advantage to late-fusion architectures over early-fusion ones, which do not rely on image encoders or tokenizers. On the contrary, early fusion exhibits stronger performance at lower parameter counts, is more efficient to train, and is easier to deploy. Motivated by the strong performance of the early-fusion architectures, we show that incorporating Mixture of Experts (MoEs) allows models to learn modality-specific weights, significantly benefiting performance.

Visual Self-Supervised Learning (SSL) currently underperforms Contrastive Language-Image Pretraining (CLIP) in multimodal settings such as Visual Question Answering (VQA). This multimodal gap is often attributed to the semantics introduced by language supervision, even though visual SSL and CLIP models are often trained on different data. In this work, we ask the question: "Do visual self-supervised approaches lag behind CLIP due to the lack of language supervision, or differences in the training data?" We study this question by training both visual SSL and CLIP models on the same MetaCLIP data, and leveraging VQA as a diverse testbed for vision encoders. In this controlled setup, visual SSL models scale better than CLIP models in terms of data and model capacity, and visual SSL performance does not saturate even after scaling up to 7B parameters. Consequently, we observe visual SSL methods achieve CLIP-level performance on a wide range of VQA and classic vision benchmarks. These findings demonstrate that pure visual SSL can match language-supervised visual pretraining at scale, opening new opportunities for vision-centric representation learning. Code and models are at https://github.com/facebookresearch/webssl.

Despite significant advancements in vision-language models (VLMs), there lacks effective approaches to enhance response quality by scaling inference-time computation. This capability is known to be a core step towards the self-improving models in recent large language model studies. In this paper, we present Vision Value Model (VisVM) that can guide VLM inference-time search to generate responses with better visual comprehension. Specifically, VisVM not only evaluates the generated sentence quality in the current search step, but also anticipates the quality of subsequent sentences that may result from the current step, thus providing a long-term value. In this way, VisVM steers VLMs away from generating sentences prone to hallucinations or insufficient detail, thereby producing higher quality responses. Experimental results demonstrate that VisVM-guided search significantly enhances VLMs' ability to generate descriptive captions with richer visual details and fewer hallucinations, compared with greedy decoding and search methods with other visual reward signals. Furthermore, we find that self-training the model with the VisVM-guided captions improve VLM's performance across a wide range of multimodal benchmarks, indicating the potential for developing self-improving VLMs.

Temporal Action Detection (TAD) is essential for analyzing long-form videos by identifying and segmenting actions within untrimmed sequences. While recent innovations like Temporal Informative Adapters (TIA) have improved resolution, memory constraints still limit large video processing. To address this, we introduce AdaTAD++, an enhanced framework that decouples temporal and spatial processing within adapters, organizing them into independently trainable modules. Our novel two-step training strategy first optimizes for high temporal and low spatial resolution, then vice versa, allowing the model to utilize both high spatial and temporal resolutions during inference, while maintaining training efficiency. Additionally, we incorporate a more sophisticated temporal module capable of capturing long-range dependencies more effectively than previous methods. Experiments on benchmark datasets, including ActivityNet-1.3, THUMOS14, and EPIC-Kitchens 100, demonstrate that AdaTAD++ achieves state-of-the-art performance. We also explore various adapter configurations, discussing their trade-offs regarding resource constraints and performance, providing valuable insights into their optimal application.

Deep learning algorithms for object classification and 3D object pose estimation lack robustness to out-of-distribution factors such as synthetic stimuli, changes in weather conditions, and partial occlusion. Recently, a class of Neural Mesh Models have been developed where objects are represented in terms of 3D meshes with learned features at the vertices. These models have shown robustness in small-scale settings, involving 10 objects, but it is unclear that they can be scaled up to 100s of object classes. The main problem is that their training involves contrastive learning among the vertices of all object classes, which scales quadratically with the number of classes. We present a strategy which exploits the compositionality of the objects, i.e. the independence of the feature vectors of the vertices, which greatly reduces the training time while also improving the performance of the algorithms. We first restructure the per-vertex contrastive learning into contrasting within class and between classes. Then we propose a process that dynamically decouples the contrast between classes which are rarely confused, and enhances the contrast between the vertices of classes that are most confused. Our large-scale 3D compositional model not only achieves state-of-the-art performance on the task of predicting classification and pose estimation simultaneously, surpassing the Neural Mesh Models and standard DNNs, but is also more robust to out-of-distribution testing including occlusion, weather conditions, synthetic data, and generalization to unknown classes.

Recent advancements in unified image generation models, such as OmniGen, have enabled the handling of diverse image generation and editing tasks within a single framework, accepting multimodal, interleaved texts and images in free form. This unified architecture eliminates the need for text encoders, greatly reducing model complexity and standardizing various image generation and editing tasks, making it more user-friendly. However, we found that it suffers from text instruction neglect, especially when the text instruction contains multiple sub-instructions. To explore this issue, we performed a perturbation analysis on the input to identify critical steps and layers. By examining the cross-attention maps of these key steps, we observed significant conflicts between neglected sub-instructions and the activations of the input image. In response, we propose Self-Adaptive Attention Scaling (SaaS), a method that leverages the consistency of cross-attention between adjacent timesteps to dynamically scale the attention activation for each sub-instruction. Our SaaS enhances instruction-following fidelity without requiring additional training or test-time optimization. Experimental results on instruction-based image editing and visual conditional image generation validate the effectiveness of our SaaS, showing superior instruction-following fidelity over existing methods. The code is available at https://github.com/zhouchao-ops/SaaS.

Direct Preference Optimization (DPO) has emerged as a powerful approach to align text-to-image (T2I) models with human feedback. Unfortunately, successful application of DPO to T2I models requires a huge amount of resources to collect and label large-scale datasets, e.g., millions of generated paired images annotated with human preferences. In addition, these human preference datasets can get outdated quickly as the rapid improvements of T2I models lead to higher quality images. In this work, we investigate a scalable approach for collecting large-scale and fully synthetic datasets for DPO training. Specifically, the preferences for paired images are generated using a pre-trained reward function, eliminating the need for involving humans in the annotation process, greatly improving the dataset collection efficiency. Moreover, we demonstrate that such datasets allow averaging predictions across multiple models and collecting ranked preferences as opposed to pairwise preferences. Furthermore, we introduce RankDPO to enhance DPO-based methods using the ranking feedback. Applying RankDPO on SDXL and SD3-Medium models with our synthetically generated preference dataset "Syn-Pic" improves both prompt-following (on benchmarks like T2I-Compbench, GenEval, and DPG-Bench) and visual quality (through user studies). This pipeline presents a practical and scalable solution to develop better preference datasets to enhance the performance of text-to-image models.

Existing vector quantization (VQ) methods struggle with scalability, largely attributed to the instability of the codebook that undergoes partial updates during training. The codebook is prone to collapse as utilization decreases, due to the progressively widening distribution gap between non-activated codes and visual features. To solve the problem, we propose Index Backpropagation Quantization (IBQ), a new VQ method for the joint optimization of all codebook embeddings and the visual encoder. Applying a straight-through estimator on the one-hot categorical distribution between the encoded feature and codebook, all codes are differentiable and maintain a consistent latent space with the visual encoder. IBQ enables scalable training of visual tokenizers and, for the first time, achieves a large-scale codebook (2^ 18 ) with high dimension (256) and high utilization. Experiments on the standard ImageNet benchmark demonstrate the scalability and superiority of IBQ, achieving competitive results on reconstruction and the application of autoregressive visual generation.

The commercialization of generative artificial intelligence (GenAI) has led to a multi-level ecosystem involving model developers, service providers, and consumers. Thus, ensuring traceability is crucial, as service providers may violate intellectual property rights (IPR), and consumers may generate harmful content. However, existing methods are limited to single-level attribution scenarios and cannot simultaneously trace across multiple levels. To this end, we introduce a scalable dual fingerprinting method for text-to-image (T2I) models, to achieve traceability of both service providers and consumers. Specifically, we propose 2-headed Fingerprint-Informed Low-Rank Adaptation (FI-LoRA), where each head is controlled by a binary fingerprint and capable of introducing the fingerprints into generated images. In practice, one FI-LoRA head is used by the developer to assign a unique fingerprint to each service provider, while the other is made available to service providers for embedding consumer-specific fingerprints during image generation. Our method does not merely embed two fingerprints within the generated image but instead allows independent control over them at developer level and business level, enabling simultaneous traceability of businesses and consumers. Experiments show that our method applies to various image generation and editing tasks of multiple T2I models, and can achieve over 99.9% extraction accuracy for both fingerprints. Our method also demonstrates good robustness against both image-level attacks and white-box model-level attacks. We hope our work provides a unified solution for developers to implement multi-tiered traceability of their models and hierarchical control over model distribution and content generation.

Recent advancements in generative models have enabled 3D urban scene generation from satellite imagery, unlocking promising applications in gaming, digital twins, and beyond. However, most existing methods rely heavily on neural rendering techniques, which hinder their ability to produce detailed 3D structures on a broader scale, largely due to the inherent structural ambiguity derived from relatively limited 2D observations. To address this challenge, we propose Sat2City, a novel framework that synergizes the representational capacity of sparse voxel grids with latent diffusion models, tailored specifically for our novel 3D city dataset. Our approach is enabled by three key components: (1) A cascaded latent diffusion framework that progressively recovers 3D city structures from satellite imagery, (2) a Re-Hash operation at its Variational Autoencoder (VAE) bottleneck to compute multi-scale feature grids for stable appearance optimization, and (3) an inverse sampling strategy enabling implicit supervision for smooth appearance transitioning. To overcome the challenge of collecting real-world city-scale 3D models with high-quality geometry and appearance, we introduce a dataset of synthesized large-scale 3D cities paired with satellite-view height maps. Validated on this dataset, our framework generates detailed 3D structures from a single satellite image, achieving superior fidelity compared to existing city generation models.

For anomaly detection (AD), early approaches often train separate models for individual classes, yielding high performance but posing challenges in scalability and resource management. Recent efforts have shifted toward training a single model capable of handling multiple classes. However, directly extending early AD methods to multi-class settings often results in degraded performance. In this paper, we investigate this performance degradation observed in reconstruction-based methods, identifying the key issue: inter-class confusion. This confusion emerges when a model trained in multi-class scenarios incorrectly reconstructs samples from one class as those of another, thereby exacerbating reconstruction errors. To this end, we propose a simple yet effective modification, called class-aware contrastive learning (CCL). By explicitly leveraging raw object category information (e.g., carpet or wood) as supervised signals, we introduce local CL to refine multiscale dense features, and global CL to obtain more compact feature representations of normal patterns, thereby effectively adapting the models to multi-class settings. Experiments across five datasets validate the effectiveness of our approach, demonstrating significant improvements and superior performance compared to state-of-the-art methods. Notably, ablation studies indicate that pseudo-class labels can achieve comparable performance.

Deep neural network (DNN)-based policy models, such as vision-language-action (VLA) models, excel at automating complex decision-making from multi-modal inputs. However, scaling these models greatly increases computational overhead, complicating deployment in resource-constrained settings like robot manipulation and autonomous driving. To address this, we propose Saliency-Aware Quantized Imitation Learning (\method), which combines quantization-aware training with a selective loss-weighting strategy for mission-critical states. By identifying these states via saliency scores and emphasizing them in the training loss, \method preserves decision fidelity under low-bit precision. We validate \method's generalization capability across extensive simulation benchmarks with environment variations, real-world tasks, and cross-domain tasks (self-driving, physics simulation), consistently recovering full-precision performance. Notably, a 4-bit weight-quantized VLA model for robotic manipulation achieves up to 2.5xspeedup and 2.5xenergy savings on an edge GPU with minimal accuracy loss. These results underline \method's potential for efficiently deploying large IL-based policy models on resource-limited devices.

Vision-Language Models (VLMs) extend the capabilities of Large Language Models (LLMs) by incorporating visual information, yet they remain vulnerable to jailbreak attacks, especially when processing noisy or corrupted images. Although existing VLMs adopt security measures during training to mitigate such attacks, vulnerabilities associated with noise-augmented visual inputs are overlooked. In this work, we identify that missing noise-augmented training causes critical security gaps: many VLMs are susceptible to even simple perturbations such as Gaussian noise. To address this challenge, we propose Robust-VLGuard, a multimodal safety dataset with aligned / misaligned image-text pairs, combined with noise-augmented fine-tuning that reduces attack success rates while preserving functionality of VLM. For stronger optimization-based visual perturbation attacks, we propose DiffPure-VLM, leveraging diffusion models to convert adversarial perturbations into Gaussian-like noise, which can be defended by VLMs with noise-augmented safety fine-tuning. Experimental results demonstrate that the distribution-shifting property of diffusion model aligns well with our fine-tuned VLMs, significantly mitigating adversarial perturbations across varying intensities. The dataset and code are available at https://github.com/JarvisUSTC/DiffPure-RobustVLM.

We present Stable Video Materials 3D (SViM3D), a framework to predict multi-view consistent physically based rendering (PBR) materials, given a single image. Recently, video diffusion models have been successfully used to reconstruct 3D objects from a single image efficiently. However, reflectance is still represented by simple material models or needs to be estimated in additional pipeline steps to enable relighting and controlled appearance edits. We extend a latent video diffusion model to output spatially-varying PBR parameters and surface normals jointly with each generated RGB view based on explicit camera control. This unique setup allows for direct relighting in a 2.5D setting, and for generating a 3D asset using our model as neural prior. We introduce various mechanisms to this pipeline that improve quality in this ill-posed setting. We show state-of-the-art relighting and novel view synthesis performance on multiple object-centric datasets. Our method generalizes to diverse image inputs, enabling the generation of relightable 3D assets useful in AR/VR, movies, games and other visual media.

Connectionist temporal classification (CTC)-based scene text recognition (STR) methods, e.g., SVTR, are widely employed in OCR applications, mainly due to their simple architecture, which only contains a visual model and a CTC-aligned linear classifier, and therefore fast inference. However, they generally exhibit worse accuracy than encoder-decoder-based methods (EDTRs) due to struggling with text irregularity and linguistic missing. To address these challenges, we propose SVTRv2, a CTC model endowed with the ability to handle text irregularities and model linguistic context. First, a multi-size resizing strategy is proposed to resize text instances to appropriate predefined sizes, effectively avoiding severe text distortion. Meanwhile, we introduce a feature rearrangement module to ensure that visual features accommodate the requirement of CTC, thus alleviating the alignment puzzle. Second, we propose a semantic guidance module. It integrates linguistic context into the visual features, allowing CTC model to leverage language information for accuracy improvement. This module can be omitted at the inference stage and would not increase the time cost. We extensively evaluate SVTRv2 in both standard and recent challenging benchmarks, where SVTRv2 is fairly compared to popular STR models across multiple scenarios, including different types of text irregularity, languages, long text, and whether employing pretraining. SVTRv2 surpasses most EDTRs across the scenarios in terms of accuracy and inference speed. Code: https://github.com/Topdu/OpenOCR.