ICCV 2025 Papers

Contemporary image generation systems have achieved high fidelity and superior aesthetic quality beyond basic text-image alignment. However, existing evaluation frameworks have failed to evolve in parallel. This study reveals that human preference reward models fine-tuned based on CLIP and BLIP architectures have inherent flaws: they inappropriately assign low scores to images with rich details and high aesthetic value, creating a significant discrepancy with actual human aesthetic preferences. To address this issue, we design a novel evaluation score, ICT (Image-Contained-Text) score, that achieves and surpasses the objectives of text-image alignment by assessing the degree to which images represent textual content. Building upon this foundation, we further train a HP (High-Preference) score model using solely the image modality, aiming to enhance image quality in aspects such as aesthetics and detail refinement while maintaining achieved text-image alignment.Experiments demonstrate that the proposed evaluation model improves scoring accuracy by over 10% compared to existing methods, and achieves significant results in optimizing state-of-the-art text-to-image models. This research provides theoretical foundation and empirical support for the evolution of image generation technology toward better alignment with higher-order human aesthetic preferences.

Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFormer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFormer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce "\text text \prec \text video " hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFormer outperforms state-of-the-art methods. Code is released at \href https://github.com/lijun2005/ICCV25-HLFormer https://github.com/lijun2005/ICCV25-HLFormer .

The optimality of using the de facto cross-entropy loss with one-hot target distribution (hard labeling) is questioned when training (Multimodal) Large Language Models (LLMs/MLLMs). Although it is reasonable for language token prediction, which is a typical multi-class classification problem in discrete space, it is suboptimal for task like numerical prediction, which is a typical regression problem in continuous space. However, enabling regression in LLMs/MLLMs will complicate the training and next-token prediction paradigm at inference. Instead, to address this challenge, we propose a novel loss design, called soft labeling, which smooths the target probability distribution, enabling predictions to be penalized according to their distance to the target. This is similar to regression loss, which penalizes more on the further predictions in the continuous space, but will not change the model architecture and the next-token prediction paradigm of LLMs/MLLMs. We demonstrate the efficacy of soft labeling through extensive experiments on visual grounding, object counting, and chart understanding, achieving state-of-the-art performance on multiple benchmarks without bells and whistles. Soft labeling can be applied in any LLM/MLLM.

Sufficient cross-task interaction is crucial for success in multi-task dense prediction. However, sufficient interaction often results in high computational complexity, forcing existing methods to face the trade-off between interaction completeness and computational efficiency. To address this limitation, this work proposes a Bidirectional Interaction Mamba (BIM), which incorporates novel scan mechanisms to adapt the Mamba modeling approach for multi-task dense prediction. On the one hand, we introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task-specific representations as bidirectional sequences during interaction. By integrating task-first and position-first scan modes within a unified linear complexity architecture, BI-Scan efficiently preserves critical cross-task information. On the other hand, we employ a Multi-Scale Scan (MS-Scan) mechanism to achieve multi-granularity scene modeling. This design not only meets the diverse granularity requirements of various tasks but also enhances nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our BIM vs its state-of-the-art competitors.

Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivariance Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization.

Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs)---a finetuning-free method that leverages sparse attention head activations (fewer than 5% of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations.

Pansharpening aims to generate high-resolution multispectral (MS) images by fusing panchromatic (PAN) images with corresponding low-resolution MS images. However, many existing methods struggle to fully capture spatial and spectral interactions, limiting their effectiveness. To address this, we propose a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging the compact representation capabilities of quaternions for high-dimensional data. Our method consists of three key components: the quaternion global spectral interaction branch, the quaternion local spatial structure awareness branch, and the quaternion spatial-spectral interaction branch. The first applies the quaternion Fourier transform to convert multi-channel features into the frequency domain as a whole, enabling global information interaction while preserving inter-channel dependencies, which aids spectral fidelity. The second uses a customized spatial quaternion representation, combined with a window-shifting strategy, to maintain local spatial dependencies while promoting spatial interactions, which helps inject spatial details. The last integrates the two pathways within the quaternion framework to enrich spatial-spectral interactions for richer representations. By utilizing quaternion's multi-dimensional representation and parameter-sharing properties, our method achieves a more compact and efficient cross-resolution, multi-band information integration, significantly improving the quality of the fused image. Extensive experiments validate the proposed method's effectiveness and its superior performance over current SOTA techniques. The code is available at https://github.com/dongli8/QuatPanNet.

Accurate stereo matching under fast motion and extreme lighting conditions is a challenge for many vision applications. Event cameras have the advantages of low latency and high dynamic range, thus providing a reliable solution to this challenge. However, since events are sparse, this makes it an ill-posed problem to obtain dense disparity using only events. In this work, we propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. Specifically, a multi-level intensity-to-event distillation strategy is designed to maximize the potential of long-range information, local texture details, and task-related knowledge of the intensity images. Simultaneously, to enforce the cross-view consistency, an intensity-event joint left-right consistency module is proposed. With our framework, extensive dense and structural information contained in intensity images is distilled to the event branch. Therefore, retaining only the events can predict dense disparities during inference, preserving the low latency characteristics of the events. Adequate experiments conducted on the MVSEC and DSEC datasets demonstrate that our method exhibits superior stereo matching performance than baselines, both quantitatively and qualitatively.

In recent years, multi-modal large language models (MLLMs) have been successfully adopted to generate humorous and engaging descriptions for internet memes. While, it is challenging for the same approaches to apply to ordinary images which lack of inherent funny or exaggerated contents. Thus, crafting appealing descriptions for ordinary image demands imaginative efforts to discover or create intriguing connections between words to image contents. To address this gap, we introduce AppealImage, a large-scale dataset consisting of ordinary images paired with appealing descriptions. AppealImage allows us to define four distinct tasks with quantitative metrics to enable objective evaluation. Subsequently, we propose CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet combines instruction tuning with heuristic active learning, guided by a referee model. Experimental results demonstrate that CharmNet outperforms the state-of-the-art method by 11.4% in generating appealing descriptions. Furthermore, CharmNet delivers impressive performance across various creative applications, including visual storytelling and situational dialogue generation. These results highlight CharmNet's potential to enhance social media engagement and to empower strong brand presence in competitive markets.

Current end-to-end multi-modal models utilize different encoders and decoders to process input and output information. This separation hinders the joint representation learning of various modalities. To unify multi-modal processing, we propose a novel architecture called MDM (Multi-modal Diffusion Mamba). MDM utilizes a Mamba-based multi-step selection diffusion model to progressively generate and refine modality-specific information through a unified variational autoencoder for both encoding and decoding. This innovative approach allows MDM to achieve superior performance when processing high-dimensional data, particularly in generating high-resolution images and extended text sequences simultaneously. Our evaluations in areas such as image generation, image captioning, visual question answering, text comprehension, and reasoning tasks demonstrate that MDM significantly outperforms existing end-to-end models (MonoFormer, LlamaGen, and Chameleon etc.) and competes effectively with SOTA models like GPT-4V, Gemini Pro, and Mistral. Our results validate MDM's effectiveness in unifying multi-modal processes while maintaining computational efficiency, establishing a new direction for end-to-end multi-modal architectures.

Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets <subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in \hyperref[URL] https://github.com/wlw951226/ARN .

End-to-end autonomous driving has achieved remarkable progress by integrating perception, prediction, and planning into a fully differentiable framework. Yet, to fully realize its potential, an effective online trajectory evaluation is indispensable to ensure safety. By forecasting the future outcomes of a given trajectory, trajectory evaluation becomes much more effective. This goal can be achieved by employing a world model to capture environmental dynamics and predict future states. Therefore, we propose an end-to-end driving framework **WoTE**, which leverages a BEV **Wo**rld model to predict future BEV states for **T**rajectory **E**valuation. The proposed BEV world model is latency-efficient compared to image-level world models and can be seamlessly supervised using off-the-shelf BEV-space traffic simulators. We validate our framework on both the NAVSIM benchmark and the closed-loop Bench2Drive benchmark based on the CARLA simulator, achieving state-of-the-art performance. Code is released at https://github.com/liyingyanUCAS/WoTE.

In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution (SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention (ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to 32x32 with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100x2 while reducing latency and memory usage by 16xand 12.2x. Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution (ESC), notably improves PSNR by 0.27 dB on Urban100x4 compared to HiT-SRF, reducing the latency and memory usage by 3.7xand 6.2x, respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module.

Recent research shows that emotions can enhance users' cognition and influence information communication. While research on visual emotion analysis is extensive, limited work has been done on helping users generate emotionally rich image content. Existing work on emotional image generation relies on discrete emotion categories, making it challenging to capture complex and subtle emotional nuances accurately. Additionally, these methods struggle to control the specific content of generated images based on text prompts. In this paper, we introduce the task of continuous emotional image content generation (C-EICG) and present EmotiCrafter, a general emotional image generation model that generates images based on free text prompts and Valence-Arousal (V-A) values. It leverages a novel emotion-embedding mapping network to fuse V-A values into textual features, enabling the capture of emotions in alignment with intended input prompts. A novel loss function is also proposed to enhance emotion expression. The experimental results show that our method effectively generates images representing specific emotions with the desired content and outperforms existing techniques.

The field of Embodied AI predominantly relies on simulation for training and evaluation, often using either fully synthetic environments that lack photorealism or high-fidelity real-world reconstructions captured with expensive hardware. As a result, sim-to-real transfer remains a major challenge. In this paper, we introduce EmbodiedSplat, a novel approach that personalizes policy training by efficiently capturing the deployment environment and fine-tuning policies within the reconstructed scenes. Our method leverages 3D Gaussian Splatting (GS) and the Habitat-Sim simulator to bridge the gap between realistic scene capture and effective training environments. Using iPhone-captured deployment scenes, we reconstruct meshes via GS, enabling training in settings that closely approximate real-world conditions. We conduct a comprehensive analysis of training strategies, pre-training datasets, and mesh reconstruction techniques, evaluating their impact on sim-to-real predictivity in real-world scenarios. Experimental results demonstrate that agents fine-tuned with EmbodiedSplat outperform both zero-shot baselines pre-trained on large-scale real-world datasets (HM3D) and synthetically generated datasets (HSSD), achieving absolute success rate improvements of 20% and 40% on real-world Image Navigation task. Moreover, our approach yields a high sim-vs-real correlation (0.87-0.97) for the reconstructed meshes, underscoring its effectiveness in adapting policies to diverse environments with minimal effort. Project page: https://gchhablani.github.io/embodied-splat.

3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOcc-ScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc.

This paper investigates the problem of understanding dynamic 3D scenes from egocentric observations, a key challenge in robotics and embodied AI. Unlike prior studies that explored this as long-form video understanding and utilized egocentric video only, we instead propose an LLM-based agent, Embodied VideoAgent, which constructs scene memory from both egocentric video and embodied sensory inputs (e.g. depth and pose sensing). We further introduce a VLM-based approach to automatically update the memory when actions or activities over objects are perceived. Embodied VideoAgent attains significant advantages over counterparts in challenging reasoning and planning tasks in 3D scenes, achieving gains of 6.5% on Ego4D-VQ3D, 2.6% on OpenEQA, and 15.3% on EnvQA. We have also demonstrated its potential in various embodied AI tasks including generating embodied interactions and perception for robot manipulation. The code and demo will be made public.