CVPR 2025 论文列表

This work presents the first condensation approach for procedural video datasets used in temporal action segmentation. We propose a condensation framework that leverages generative prior learned from the dataset and network inversion to condense data into compact latent codes with significant storage reduced across temporal and channel aspects. Orthogonally, we propose sampling diverse and representative action sequences to minimize video-wise redundancy. Our evaluation on standard benchmarks demonstrates consistent effectiveness in condensing TAS datasets and achieving competitive performances. Specifically, on the Breakfast dataset, our approach reduces storage by over 500xwhile retaining 83% of the performance compared to training with the full dataset. Furthermore, when applied to a downstream incremental learning task, it yields superior performance compared to the state-of-the-art.

Diffusion customization methods have achieved impressive results with only a minimal number of user-provided images. However, existing approaches customize concepts collectively, whereas real-world applications often require sequential concept integration. This sequential nature can lead to catastrophic forgetting, where previously learned concepts are lost. In this paper, we investigate concept forgetting and concept confusion in the continual customization. To tackle these challenges, we present ConceptGuard, a comprehensive approach that combines shift embedding, concept-binding prompts and memory preservation regularization, supplemented by a priority queue which can adaptively update the importance and occurrence order of different concepts. These strategies can dynamically update, unbind and learn the relationship of the previous concepts, thus alleviating concept forgetting and confusion. Through comprehensive experiments, we show that our approach outperforms all the baseline methods consistently and significantly in both quantitative and qualitative analyses.

As large-scale diffusion models continue to advance, they excel at producing high-quality images but often generate unwanted content, such as sexually explicit or violent content. Existing methods for concept removal generally guide the image generation process but can unintentionally modify unrelated regions, leading to inconsistencies with the original model. We propose a novel approach for targeted concept replacing in diffusion models, enabling specific concepts to be removed without affecting non-target areas. Our method introduces a dedicated concept localizer for precisely identifying the target concept during the denoising process, trained with few-shot learning to require minimal labeled data. Within the identified region, we introduce a training-free Dual Prompts Cross-Attention (DPCA) module to substitute the target concept, ensuring minimal disruption to surrounding content. We evaluate our method on concept localization precision and replacement efficiency. Experimental results demonstrate that our method achieves superior precision in localizing target concepts and performs coherent concept replacement with minimal impact on non-target areas, outperforming existing approaches.

Diffusion models are widely used for image editing tasks. Existing editing methods often design a representation manipulation procedure by curating an edit direction in the text embedding or score space. However, such a procedure faces a key challenge: overestimating the edit strength harms visual consistency while underestimating it fails the editing task. Notably, each source image may require a different editing strength, and it is costly to search for an appropriate strength via trial-and-error. To address this challenge, we propose Concept Lancet (CoLan), a zero-shot plug-and-play framework for principled representation manipulation in diffusion-based image editing. At inference time, we decompose the source input in the latent (text embedding or diffusion score) space as a sparse linear combination of the representations of the collected visual concepts and phrases. This allows us to accurately estimate the presence of concepts in each image, which informs the edit. Based on the editing task (replace/add/remove), we perform a customized concept transplant process to impose the corresponding editing direction. To sufficiently model the concept space, we curate a conceptual representation dataset, CoLan-150K, which contains diverse descriptions and scenarios of visual concepts and phrases for the latent dictionary. Experiments on multiple diffusion-based image editing baselines show that methods equipped with CoLan achieve state-of-the-art performance in editing effectiveness and consistency preservation.

Composed image retrieval (CIR) is the task of retrieving a target image specified by a query image and a relative text that describes a semantic modification to the query image. Existing methods in CIR struggle to accurately represent the image and the text modification, resulting in subpar performance. To address this limitation, we introduce a CIR framework, \name , trained with a Text Concept-Consistency loss that encourages the representations of noun phrases in the text modification to better attend to the relevant parts of the query image. To support training with this loss function, we also propose a synthetic data generation pipeline that creates training data from existing CIR datasets or unlabeled images. We show that these components together enable stronger performance on CIR tasks, setting a new state-of-the-art in composed image retrieval in both the supervised and zero-shot settings on multiple benchmark datasets, including CIRR and CIRCO. Source code, model checkpoints, and our new datasets are available at https://github.com/mvrl/ConText-CIR.

The development of Text-to-Video (T2V) generation has made motion transfer possible, enabling the control of video motion based on existing footage. However, current methods have two limitations: 1) struggle to handle multi-subjects videos, failing to transfer specific subject motion; 2) struggle to preserve the diversity and accuracy of motion as transferring to subjects with varying shapes. To overcome these, we introduce ConMo, a zero-shot framework that disentangle and recompose the motions of subjects and camera movements. ConMo isolates individual subject and background motion cues from complex trajectories in source videos using only subject masks, and reassembles them for target video generation. This approach enables more accurate motion control across diverse subjects and improves performance in multi-subject scenarios. Additionally, we propose soft guidance in the recomposition stage which controls the retention of original motion to adjust shape constraints, aiding subject shape adaptation and semantic transformation. Unlike previous methods, ConMo unlocks a wide range of applications, including subject size and position editing, subject removal, semantic modifications, and camera motion simulation. Extensive experiments demonstrate that ConMo significantly outperforms state-of-the-art methods in motion fidelity and semantic consistency. The code is available at https://github.com/Andyplus1/ConMo.

This paper introduces Comprehensive Relighting, the first all-in-one approach that can both control and harmonize the lighting from an image or video of humans with arbitrary body parts from any scene. Building such a generalizable model is extremely challenging due to the lack of dataset, restricting existing image-based relighting models to a specific scenario (e.g., face or static human). To address this challenge, we repurpose a pre-trained diffusion model as a general image prior and jointly model the human relighting and background harmonization in the coarse-to-fine framework. To further enhance the temporal coherence of the relighting, we introduce an unsupervised temporal lighting model that learns the lighting cycle consistency from many real-world videos without any ground truth. In inference time, our temporal lighting module is combined with the diffusion models through the spatio-temporal feature blending algorithms without extra training; and we apply a new guided refinement as a post-processing to preserve the high-frequency details from the input image. In the experiments, Comprehensive Relighting shows a strong generalizability and lighting temporal coherence, outperforming existing image-based human relighting and harmonization methods.

Generating targeted universal perturbations for multi-label recognition is a combinatorially hard problem that requires exponential time and space complexity. To address the problem, we propose a compositional framework. We show that a simple independence assumption on label-wise universal perturbations naturally leads to an efficient optimization that requires learning affine convex cones spanned by label-wise universal perturbations, significantly reducing the problem complexity to linear time and space. During inference, the framework allows generating universal perturbations for novel combinations of classes in constant time. We demonstrate the scalability of our method on large datasets and target sizes, evaluating its performance on NUS-WIDE, MS-COCO, and OpenImages using state-of-the-art multi-label recognition models. Our results show that our approach outperforms baselines and achieves results comparable to methods with exponential complexity.

Image composition and generation are processes where the artists need control over various parts of the generated images. However, the current state-of-the-art generation models, like Stable Diffusion, cannot handle fine-grained part-level attributes in the text prompts. Specifically, when additional attribute details are added to the base text prompt, these text-to-image models either generate an image vastly different from the image generated from the base prompt or ignore the attribute details. To mitigate these issues, we introduce PartComposer, a training-free method that enables image generation based on fine-grained part-level attributes specified for objects in the base text prompt. This allows more control for artists and enables novel object compositions by combining distinctive object parts. PartComposer first localizes object parts by denoising the object region from a specific diffusion process. This enables each part token to be localized to the right region. After obtaining part masks, we run a localized diffusion process in each part region based on fine-grained part attributes and combine them to produce the final image. All stages of PartComposer are based on repurposing a pre-trained diffusion model, which enables it to generalize across domains. We demonstrate the effectiveness of part-level control provided by PartComposer through qualitative visual examples and quantitative comparisons with contemporary baselines.

Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference.We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce "complexity experts" -- flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity.To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at \href https://eduardzamfir.github.io/moceir/ \texttt eduardzamfir.github.io/MoCE-IR/

In this paper, we introduce the Selective Image Guided Network (SigNet), a novel degradation-aware framework that transforms depth completion into depth enhancement for the first time. Moving beyond direct completion using convolutional neural networks (CNNs), SigNet initially densifies sparse depth data through non-CNN densification tools to obtain coarse yet dense depth. This approach eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, SigNet redefines completion as enhancement, establishing a self-supervised degradation bridge between the coarse depth and the targeted dense depth for effective RGB-D fusion. To achieve this, SigNet leverages the implicit degradation to adaptively select high-frequency components (e.g., edges) of RGB data to compensate for the coarse depth. This degradation is further integrated into a multi-modal conditional Mamba, dynamically generating the state parameters to enable efficient global high-frequency information interaction. We conduct extensive experiments on the NYUv2, DIML, SUN RGBD, and TOFDC datasets, demonstrating the state-of-the-art (SOTA) performance of SigNet.

Existing single-image denoising algorithms often struggle to restore details when dealing with complex noisy images. The introduction of near-infrared (NIR) images offers new possibilities for RGB image denoising. However, due to the inconsistency between NIR and RGB images, the existing works still struggle to balance the contributions of two fields in the process of image fusion. In response to this, in this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We first propose the frequency correlation prior based on an in-depth statistical frequency analysis of NIR-RGB image pairs. The prior reveals the complementary correlation of NIR and RGB images in the frequency domain. Leveraging frequency correlation prior, we then establish a frequency learning framework composed of Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). FDSM dynamically selects complementary information from NIR and RGB images in the frequency domain, and FEFM strengthens the control of common and differential features during the fusion process of NIR and RGB features. Extensive experiments on simulated and real data validate that the proposed method outperforms other state-of-the-art methods. The code will be released at https://github.com/yuchenwang815/FCENet.

Existing approaches for controlling text-to-image diffusion models, while powerful, do not allow for explicit 3D object-centric control, such as precise control of object orientation. In this work, we address the problem of multi-object orientation control in text-to-image diffusion models. This enables the generation of diverse multi-object scenes with precise orientation control for each object. The key idea is to condition the diffusion model with a set of orientation-aware compass tokens, one for each object, along with text tokens. A light-weight encoder network predicts these compass tokens taking object orientation as the input. The model is trained on a synthetic dataset of procedurally generated scenes, each containing one or two 3D-assets on a plain background. However, direct training this framework results in poor orientation control as well as leads to entanglement among objects. To mitigate this, we intervene in the generation process and constrain the cross-attention maps of each compass token to its corresponding object regions. The trained model is able to achieve precise orientation control for a) complex objects not seen during training, and b) multi-object scenes with more than two objects, indicating strong generalization capabilities. Further, when combined with personalization methods, our method precisely controls the orientation of the new object in diverse contexts. Our method achieves state-of-the-art orientation control and text alignment quantified with extensive evaluations and a user study.

Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces CompGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed: (1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. CompGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, CompGS optimizes across objects of varying scales by dynamically adjusting the spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on T3Bench demonstrate the effectiveness of CompGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. CompGS can also be easily extended to progressively adding 3D objects, facilitating complex scene generation. We hope CompGS will provide new insights to the compositional 3D generation.

One of the key challenges of detecting AI-generated images is spotting images that have been created by previously unseen generative models. We argue that the limited diversity of the training data is a major obstacle to addressing this problem, and we propose a new dataset that is significantly larger and more diverse than prior works. As part of creating this dataset, we systematically download thousands of text-to-image latent diffusion models and sample images from them. We also collect images from dozens of popular open source and commercial models. The resulting dataset contains 2.7M images that have been sampled from 4803 different models. These images collectively capture a wide range of scene content, generator architectures, and image processing settings. Using this dataset, we study the generalization abilities of fake image detectors. Our experiments suggest that detection performance improves as the number of models in the training set increases, even when these models have similar architectures. We also find that increasing the diversity of the models improves detection performance, and that our trained detectors generalize better than those trained on other datasets. The dataset can be found in https://jespark.net/projects/2024/community_forensics

This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video- and image-based VLMs across reasoning types.To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrite VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types.

3D morphable models (3DMMs) are a powerful tool to represent the possible shapes and appearances of an object category. Given a single test image, 3DMMs can be used to solve various tasks, such as predicting the 3D shape, pose, semantic correspondence, and instance segmentation of an object. Unfortunately, 3DMMs are only available for very few object categories that are of particular interest, like faces or human bodies, as they require a demanding 3D data acquisition and category-specific training process. In contrast, we introduce a new method, Common3D, that learns 3DMMs of common objects in a fully self-supervised manner from a collection of object-centric videos. For this purpose, our model represents objects as a learned 3D template mesh and a deformation field that is parameterized as an image-conditioned neural network. Different from prior works, Common3D represents the object appearance with neural features instead of RGB colors, which enables the learning of more generalizable representations through an abstraction from pixel intensities. Importantly, we train the appearance features using a contrastive objective by exploiting the correspondences defined through the deformable template mesh. This leads to higher quality correspondence features compared to related works and a significantly improved model performance at estimating 3D object pose and semantic correspondence. Common3D is the first completely self-supervised method that can solve various vision tasks in a zero-shot manner. github.com/GenIntel/common3d.

Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents's ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only 15% of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems. Our code is available at: https://github.com/xxyQwQ/ComfyBench.

The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by 1.6% at training resolution and 2.9% at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE