ICCV 2025 论文列表

In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead.

This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal rotary position embedding to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties-including scalability, cross-modal information flow patterns, and visual representation capabilities-with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT-22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL.

The fine-tuning of large vision-language foundation models remains an underexplored area, particularly regarding its impact on learning gains and catastrophic forgetting. Inspired by the significance of modality gaps in contrastive dual-encoders, we introduce the Inter-Intra Modal Measure (IIMM)--a predictive metric that quantifies the relationship between intra-modal image embedding similarity and inter-modal misalignment. Through extensive empirical analysis across four state-of-the-art vision-language models and five fine-tuning techniques, we establish a strong linear relationship: tasks with higher IIMM scores yield greater in-domain performance improvements but suffer from more pronounced out-of-domain degradation, with some parameter-efficient fine-tuning (PEFT) methods exhibiting severe forgetting. Compared to existing transferability measures, the IIMM demonstrates significantly stronger predictive power for accuracy changes post fine-tuning in dual-encoder models. Moreover, we provide a theoretical bound, proving that changes in IIMM are limited by the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring its stability and robustness as a predictive measure. With only a single forward pass of the target data, practitioners can leverage this key insight to evaluate the degree to which a model can be expected to improve following fine-tuning. When combined with prior knowledge of a model's performance across diverse tasks, the IIMM further enhances transferability predictions for novel tasks, offering a lightweight yet effective tool for guiding model adaptation strategies. Our code is provided at https://github.com/mit-ll/IIMM.

Given a textual query along with a corresponding video, the objective of moment retrieval aims to localize the moments relevant to the query within the video. While commendable results have been demonstrated by existing transformer-based approaches, predicting the accurate temporal span of the target moment is still a major challenge. This paper reveals that a crucial reason stems from the spurious correlation between the text query and the moment context. Namely, the model makes predictions by overly associating queries with background frames rather than distinguishing target moments. To address this issue, we propose a dynamic learning approach for moment retrieval, where two strategies are designed to mitigate the spurious correlation. First, we introduce a novel video synthesis approach to construct a dynamic context for the queried moment, enabling the model to attend to the target moment of the corresponding query across dynamic backgrounds. Second, to alleviate the over-association with backgrounds, we enhance representations temporally by incorporating text-dynamics interaction, which encourages the model to align text with target moments through complementary dynamic representations. With the proposed method, our model significantly alleviates the spurious correlation issue in moment retrieval and establishes new state-of-the-art performance on two popular benchmarks, i.e., QVHighlights and Charades-STA. In addition, detailed ablation studies and evaluations across different architectures demonstrate the generalization and effectiveness of the proposed strategies. Our code is available at https://github.com/xyangzhou/TD-DETR.

Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport (C2OT) that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians-to-moons, CIFAR-10, ImageNet-32x32, and ImageNet-256x256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at hkchengrex.github.io/C2OT

Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving a ~14,000xcompression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBench T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/.

Textured 3D morphing creates smooth and plausible interpolation sequences between two 3D objects, focusing on transitions in both shape and texture. This is important for creative applications like visual effects in filmmaking. Previous methods rely on establishing point-to-point correspondences and determining smooth deformation trajectories, which inherently restrict them to shape-only morphing on untextured, topologically aligned datasets. This restriction leads to labor-intensive preprocessing and poor generalization. To overcome these challenges, we propose a method for 3D regenerative morphing using a 3D diffusion prior. Unlike previous methods that depend on explicit correspondences and deformations, our method eliminates the additional need for obtaining correspondence and uses the 3D diffusion prior to generate morphing. Specifically, we first introduce a 3D diffusion model and interpolate the source and target information at three levels: initial noise, model parameters, and condition features. We then explore an Attention Fusion strategy to generate smoother morphing sequences. To further improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering, where we match approximate tokens based on semantic analysis to guide implicit correspondences in the denoising process of the diffusion model, and (b) Low-Frequency Enhancement, where we enhance low-frequency signals in the tokens to improve the quality of generated surfaces. Experimental results show that our method achieves superior smoothness and plausibility in 3D morphing across diverse cross-category object pairs, offering a novel regenerative method for 3D morphing with textured representations.

Scene text recognition (STR) suffers from challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained models. Meanwhile, despite producing holistically appealing text images, diffusion-based visual text generation methods struggle to synthesize accurate and realistic instance-level text at scale. To tackle this, we introduce TextSSR: a novel pipeline for Synthesizing Scene Text Recognition training data. TextSSR targets three key synthesizing characteristics: accuracy, realism, and scalability. It achieves accuracy through a proposed region-centric text generation with position-glyph enhancement, ensuring proper character placement. It maintains realism by guiding style and appearance generation using contextual hints from surrounding text or background. This character-aware diffusion architecture enjoys precise character-level control and semantic coherence preservation, without relying on natural language prompts. Therefore, TextSSR supports large-scale generation through combinatorial text permutations. Based on these, we present TextSSR-F, a dataset of 3.55 million quality-screened text instances. Extensive experiments show that STR models trained on TextSSR-F outperform those trained on existing synthetic datasets by clear margins on common benchmarks, and further improvements are observed when mixed with real-world training data. Code is available at https://github.com/YesianRohn/TextSSR.

In image editing tasks, high-quality text editing capabilities can significantly reduce both human and material resource costs. Existing methods, however, face significant limitations in terms of stroke accuracy for complex text and controllability of generated text styles. To address these challenges, we propose TextMaster, a solution capable of accurately editing text across various scenarios and image regions, while ensuring proper layout and controllable text style. Our method enhances the accuracy and fidelity of text rendering by incorporating high-resolution standard glyph information and applying perceptual loss within the text editing region. Additionally, we leverage an attention mechanism to compute intermediate layer bounding box regression loss for each character, enabling the model to learn text layout across varying contexts. Furthermore, we propose a novel style injection technique that enables controllable style transfer for the injected text. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method.

Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures.Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem.These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images.This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS).The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation.We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a more accurate target distribution and semantic guidance.Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures.Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling.

Existing outfit recommendation frameworks focus on outfit compatibility prediction and complementary item retrieval. We present a text-driven outfit generation framework, Text2Outfit, which generates outfits controlled by text prompts. Our framework supports two forms of outfit recommendation: 1) Text-to-outfit generation, where the prompt includes the specification for each outfit item (e.g., product features), and the model retrieves items that match the prompt and are stylistically compatible. 2) Seed-to-outfit generation, where the prompt includes the specification for a seed item, and the model both predicts which product types the outfit should include (referred to as composition generation) and retrieves the remaining items to build an outfit. We develop a large language model (LLM) framework that learns the cross-modal mapping between text and image set, and predicts a set of embeddings and compositions to retrieve outfit items. We devise an attention masking mechanism in LLM to handle the alignment between text descriptions and image tokens from different categories. We conduct experiments on the Polyvore dataset and evaluate the quality of the generated outfits from two perspectives: 1) feature matching for outfit items, and 2) outfit visual compatibility. The results demonstrate that our approach significantly outperforms the baseline methods in text to outfit generation.

Recent advances in text-driven motion generation have shown notable advancements. However, these works are typically limited to standardized skeletons and rely on a cumbersome retargeting process to adapt to varying skeletal configurations of diverse characters. In this paper, we present OmniSkel, a novel framework that can directly generate high-quality human motions for any user-defined skeleton without retargeting. Specifically, we introduce skeleton-aware RVQ-VAE, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Moreover, we propose a simple yet effective training-free approach, Motion Restoration Optimizer (MRO), to ensure zero bone length error while preserving motion smoothness. To facilitate our research, we construct SkeleMotion-3D, a large-scale text-skeleton-motion dataset based on HumanML3D. Extensive experiments demonstrate the excellent robustness and generalization of our method.

Recent advancements in multimodal vision models have highlighted limitations in late-stage feature fusion and suboptimal query selection for hybrid prompts open-world segmentation, alongside constraints from caption-derived vocabularies. To address these challenges, we propose Prompt-DINO, a text-guided visual Prompt DINO framework featuring three key innovations. First, we introduce an early fusion mechanism that unifies text/visual prompts and backbone features at the initial encoding stage, enabling deeper cross-modal interactions to resolve semantic ambiguities. Second, we design order-aligned query selection for DETR-based architectures, explicitly optimizing the structural alignment between text and visual queries during decoding to enhance semantic-spatial consistency. Third, we develop a generative data engine powered by the Recognize Anything via Prompting (RAP) model, which synthesizes 0.5B diverse training instances through a dual-path cross-verification pipeline, reducing label noise by 80.5% compared to conventional approaches. Extensive experiments demonstrate that Prompt-DINO achieves state-of-the-art performance on open-world detection benchmarks while significantly expanding semantic coverage beyond fixed-vocabulary constraints. Our work establishes a new paradigm for scalable multimodal detection and data generation in open-world scenarios. Data&Code are available at https://github.com/WeChatCV/WeVisionOne.

Infrared small target detection is currently a hot and challenging task in computer vision. Existing methods usually focus on mining visual features of targets, which struggles to cope with complex and diverse detection scenarios. The main reason is that infrared small targets have limited image information on their own, thus relying only on visual features fails to discriminate targets and interferences, leading to lower detection performance. To address this issue, we introduce a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new research idea. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) to facilitate information fusion between texts and images. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. Moreover, proposed Text-IRSTD shows strong generalization and wide application prospects in unseen detection scenarios. The dataset and code will be publicly released after acceptance of this paper. The dataset and code are available at https://github.com/Zhengsy0407/Text-IRSTD.

Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets.

Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about 3% Dice score improvements across three datasets while reducing computational complexity by over 7 times.

Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA.

Zero-shot depth completion with metric scales poses significant challenges, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet these efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which fine-tunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea of this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Our source code is available in the supplementary materials.

We present TerraMind, the first any-to-any generative, multi-modal foundation model for Earth observation (EO). Unlike other multimodal models, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind leverages fine-grained representations to capture critical spatial nuances. We pretrained TerraMind on nine geospatial modalities of a global, large-scale dataset. In this paper, we demonstrate that (i) TerraMind's dual-scale early fusion approach unlocks a range of zero-shot and few-shot applications for Earth observation, (ii) TerraMind introduces "thinking in modalities" (TiM)--the capability of generating additional artificial data during finetuning and inference to improve the model output--and (iii) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks for EO like PANGAEA. All models and code have been open-sourced under a permissive license at https://huggingface.co/ibm-esa-geospatial and https://github.com/ibm/terramind.