CVPR 2025 Papers

Heterogeneous Federated Learning (HFL) has received widespread attention due to its adaptability to different models and data. The HFL approach utilizing auxiliary models for knowledge transfer enhances flexibility. However, existing frameworks face the challenges of aggregation bias and local overfitting. To address these issues, we propose FedSCE. It reduces the degree of freedom of update and improves generalization performance by limiting the specific layer of local model update to the local subspace. The subspace is dynamically updated to ensure coverage of the latest model update trajectory. Additionally, FedSCE evaluates client contributions based on the update distance of the auxiliary model in feature space and parameter space, achieving adaptive weighted aggregation. We validate our approach in both feature-skewed and label-skewed scenarios, demonstrating that on Office10, our method exceeds the best baseline by 3.87%. The code will be available at https://github.com/AVC2-UESTC/FedSCE.git.

N-shot neural architecture search (NAS) exploits a supernet containing all candidate subnets for a given search space. The subnets are typically trained with a static training strategy (e.g., using the same learning rate (LR) scheduler and optimizer for all subnets). This, however, does not consider that individual subnets have distinct characteristics, leading to two problems: (1) The supernet training is biased towards the low-complexity subnets (unfairness); (2) the momentum update in the supernet is noisy (noisy momentum). We present a dynamic supernet training technique to address these problems by adjusting the training strategy adaptive to the subnets. Specifically, we introduce a complexity-aware LR scheduler (CaLR) that controls the decay ratio of LR adaptive to the complexities of subnets, which alleviates the unfairness problem. We also present a momentum separation technique (MS). It groups the subnets with similar structural characteristics and uses a separate momentum for each group, avoiding the noisy momentum problem. Our approach can be applicable to various N-shot NAS methods with marginal cost, while improving the search performance drastically. We validate the effectiveness of our approach on various search spaces (e.g., NAS-Bench-201, Mobilenet spaces) and datasets (e.g., CIFAR-10/100, ImageNet).

Text-driven style transfer aims to merge the style of a reference image with content described by a text prompt. Recent advancements in text-to-image models have improved the nuance of style transformations, yet significant challenges remain, particularly with overfitting to reference styles, limiting stylistic control, and misaligning with textual content.In this paper, we propose three complementary strategies to address these issues. First, we introduce a cross-modal Adaptive Instance Normalization (AdaIN) mechanism for better integration of style and text features, enhancing alignment. Second, we develop a Style-based Classifier-Free Guidance (SCFG) approach that enables selective control over stylistic elements, reducing irrelevant influences. Finally, we incorporate a teacher model during early generation stages to stabilize spatial layouts and mitigate artifacts. Our extensive evaluations demonstrate significant improvements in style transfer quality and alignment with textual prompts. Furthermore, our approach can be integrated into existing style transfer frameworks without fine-tuning.

Training-free diffusion-based methods have achieved remarkable success in style transfer, eliminating the need for extensive training or fine-tuning. However, due to the lack of targeted training for style information extraction and constraints on the content image layout, training-free methods often suffer from layout changes of original content and content leakage from style images. Through a series of experiments, we discovered that an effective startpoint in the sampling stage significantly enhances the style transfer process. Based on this discovery, we propose StyleSSP, which focuses on obtaining a better startpoint to address layout changes of original content and content leakage from style image. StyleSSP comprises two key components: (1) Frequency Manipulation: To improve content preservation, we reduce the low-frequency components of the DDIM latent, allowing the sampling stage to pay more attention to the layout of content images; and (2) Negative Guidance via Inversion: To mitigate the content leakage from style image, we employ negative guidance in the inversion stage to ensure that the startpoint of the sampling stage is distanced from the content of style image. Experiments show that StyleSSP surpasses previous training-free style transfer baselines, particularly in preserving original content and minimizing the content leakage from style image.

Style control has been popular in video generation models. Existing methods often generate videos far from the given style, cause content leakage, and struggle to transfer one video to the desired style. Our first observation is that the style extraction stage matters, whereas existing methods emphasize global style but ignore local textures. In order to bring texture features while preventing content leakage, we filter content-related patches while retaining style ones based on prompt-patch similarity; for global style extraction, we generate a paired style dataset through model illusion to facilitate contrastive learning, which greatly enhances the absolute style consistency. Moreover, to fill in the image-to-video gap, we train a lightweight motion adapter on still videos, which implicitly enhances stylization extent, and enables our image-trained model to be seamlessly applied to videos. Benefited from these efforts, our approach, StyleMaster, not only achieves significant improvement in both style resemblance and temporal coherence, but also can easily generalize to video style transfer with a gray tile ControlNet. Extensive experiments and visualizations demonstrate that StyleMaster significantly outperforms competitors, effectively generating high-quality stylized videos that align with textual content and closely resemble the style of reference images.

We present Text-driven object-centric style editing model named Style-Editor, a novel method that guides style editing at an object-centric level using textual inputs.The core of Style-Editor is our Patch-wise Co-Directional (PCD) loss, meticulously designed for precise object-centric editing that are closely aligned with the input text. This loss combines a patch directional loss for text-guided style direction and a patch distribution consistency loss for even CLIP embedding distribution across object regions. It ensures a seamless and harmonious style editing across object regions.Key to our method are the Text-Matched Patch Selection (TMPS) and Pre-fixed Region Selection (PRS) modules for identifying object locations via text, eliminating the need for segmentation masks. Lastly, we introduce an Adaptive Background Preservation (ABP) loss to maintain the original style and structural essence of the image's background. This loss is applied to dynamically identified background areas.Extensive experiments underline the effectiveness of our approach in creating visually coherent and textually aligned style editing.

Under limited data setting, GANs often struggle to navigate and effectively exploit the input latent space. Consequently, images generated from adjacent variables in a sparse input latent space may exhibit significant discrepancies in realism, leading to suboptimal consistency regularization (CR) outcomes. To address this, we propose SQ-GAN, a novel approach that enhances CR by introducing a style space quantization scheme. This method transforms the sparse, continuous input latent space into a compact, structured discrete proxy space, allowing each element to correspond to a specific data point, thereby improving CR performance. Instead of direct quantization, we first map the input latent variables into a less entangled "style" space and apply quantization using a learnable codebook. This enables each quantized code to control distinct factors of variation. Additionally, we optimize the optimal transport distance to align the codebook codes with features extracted from the training data by a foundation model, embedding external knowledge into the codebook and establishing a semantically rich vocabulary that properly describes the training dataset. Extensive experiments demonstrate significant improvements in both discriminator robustness and generation quality with our method.

Recently, a task of Single-Domain Generalized Object Detection (Single-DGOD) is proposed, aiming to generalize a detector to multiple unknown domains never seen before during training. Due to the unavailability of target-domain data, some methods leverage the multimodal capabilities of vision-language models, using textual prompts to estimate cross-domain information, enhancing the model's generalization capability. These methods typically use a single textual prompt, referred to as the one-step prompt method. However, when dealing with complex styles, such as the combination of rain and night, we observe that the performance of the one-step prompt method tends to be relatively weak. The reason may be that many scenes incorporate a single style and a combination of multiple styles. The one-step prompt method may not effectively synthesize combined information involving various styles. To address this limitation, we propose a new method, i.e., Style Evolving along Chain-of-Thought, which aims to progressively integrate and expand style information along the chain of thought, enabling the continual evolution of styles. Specifically, by progressively refining style descriptions and guiding the diverse evolution of styles, this method enhances the simulation of various style characteristics, enabling the model to learn and adapt to subtle differences more effectively. Additionally, it exposes the model to a broader range of style features with different data distributions, thereby enhancing its generalization capability in unseen domains. The significant performance gains over five adverse-weather scenarios and the Real to Art benchmark demonstrate the superiorities of our method. Our code is available at https://github.com/ZZ2490/SE-COT.

We introduce a novel 3D generation method for versatile and high-quality 3D asset creation.The cornerstone is a unified Structured LATent (SLAT) representation which allows decoding to different output formats, such as Radiance Fields, 3D Gaussians, and meshes. This is achieved by integrating a sparsely-populated 3D grid with dense multiview visual features extracted from a powerful vision foundation model, comprehensively capturing both structural (geometry) and textural (appearance) information while maintaining flexibility during decoding.We employ rectified flow transformers tailored for SLAT as our 3D generation models and train models with up to 2 billion parameters on a large 3D asset dataset of 500K diverse objects. Our model generates high-quality results with text or image conditions, significantly surpassing existing methods, including recent ones at similar scales. We showcase flexible output format selection and local 3D editing capabilities which were not offered by previous models. Code, model, and data will be released.

In this paper, we present a new generic Structure-from-Motion pipeline, GenSfM, that uses a non-parametric camera projection model. The model is self-calibrated during the reconstruction process and can fit a wide variety of cameras, ranging from simple low-distortion pinhole cameras to more extreme optical systems such as fisheye or catadioptric cameras. The key component in our framework is an adaptive calibration procedure that can estimate partial calibrations, only modeling regions of the image where sufficient constraints are available. In experiments, we show that our method achieves comparable accuracy to traditional Structure-from-Motion pipelines in easy scenarios, and outperforms them in cases where they are unable to self-calibrate their parametric models. Code is at https://github.com/Ivonne320/GenSfM.git

Relative pose estimation provides a promising way for achieving object-agnostic pose estimation. Despite the success of existing 3D correspondence-based methods, the reliance on explicit feature matching suffers from small overlaps in visible regions and unreliable feature estimation for invisible regions. Inspired by humans' ability to assemble two object parts that have small or no overlapping regions by considering object structure, we propose a novel Structure-Aware Correspondence Learning method for Relative Pose Estimation, which consists of two key modules. First, a structure-aware keypoint extraction module is designed to locate a set of kepoints that can represent the structure of objects with different shapes and appearance, under the guidance of a keypoint based image reconstruction loss. Second, a structure-aware correspondence estimation module is designed to model the intra-image and inter-image relationships between keypoints to extract structure-aware features for correspondence estimation. By jointly leveraging these two modules, the proposed method can naturally estimate 3D-3D correspondences for unseen objects without explicit feature matching for precise relative pose estimation. Experimental results on the CO3D, Objaverse and LineMOD datasets demonstrate that the proposed method significantly outperforms prior methods, i.e., with 5.7^\circ reduction in mean angular error on the CO3D dataset.

Recent advancements in neural 3D representations, such as neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS), have made accurate estimation of the 3D structure from multiview images possible. However, this capability is limited to estimating the visible external structure, and it is still difficult to identify the invisible internal structure hidden behind the surface. To overcome this limitation, we address a new task called structure from collision (SfC), which aims to estimate the structure (including the invisible internal one) of an object from the appearance changes at collision. To solve this task, we propose a novel model called SfC-NeRF, which optimizes the invisible internal structure of the object through a video sequence under physical, appearance (i.e., visible external structure)-preserving, and keyframe constraints. In particular, to avoid falling into undesirable local optima owing to its ill-posed nature, we propose volume annealing, i.e., searching for the global optima by repeatedly reducing and expanding the volume. Extensive experiments on 115 objects involving diverse structures (i.e., various cavity shapes, locations, and sizes) and various material properties reveal the properties of SfC and demonstrate the effectiveness of the proposed SfC-NeRF.

As scaling laws in generative AI push performance, they simultaneously concentrate the development of these models among actors with large computational resources. With a focus on text-to-image (T2I) generative models, we aim to unlock this bottleneck by demonstrating very low-cost training of large-scale T2I diffusion transformer models. As the computational cost of transformers increases with the number of patches in each image, we propose randomly masking up to 75% of the image patches during training. We propose a deferred masking strategy that preprocesses all patches using a patch-mixer before masking, thus significantly reducing the performance degradation with masking, making it superior to model downscaling in reducing computational cost. We also incorporate the latest improvements in transformer architecture, such as the use of mixture-of-experts layers, to improve performance and further identify the critical benefit of using synthetic images in micro-budget training. Finally, using only 37M publicly available real and synthetic images, we train a 1.16 billion parameter sparse transformer with only 1,890 USD economical cost and achieve a 12.7 FID in zero-shot generation on the COCO dataset. Notably, our model achieves competitive performance across both automated and human-centric evaluations, as well as high-quality generations, while incurring 118xlower costs than Stable Diffusion models and 14xlower costs than the current state-of-the-art approach, which costs \28,400. We also further investigate the influence of synthetic images on performance and demonstrate that micro-budget training on only synthetic images is sufficient for achieving high-quality data generation. Our end-to-end training pipeline and model checkpoints are available at https://github.com/SonyResearch/micro_diffusion.

This paper aims to tackle the problem of photorealistic view synthesis from vehicle sensors data. Recent advancements in neural scene representation have achieved notable success in rendering high-quality autonomous driving scenes,but the performance significantly degrades as the viewpoint deviates from the training trajectory. To mitigate this problem, we introduce StreetCrafter, a novel controllable video diffusion model that utilizes LiDAR point cloud renderings as pixel-level conditions, which fully exploits the generative prior for novel view synthesis, while preserving precise camera control.Moreover, the utilization of pixel-level LiDAR condition allows us to make accurate pixel-level edits to target scenes.In addition, the generative prior of StreetCrafter can be effectively incorporated into dynamic scene representations to achieve real-time rendering. Experiments on Waymo Open and Pandaset datasets demonstrate that our model enables flexible control over viewpoint changes, enlarging the view synthesis regions for satisfying rendering, which outperforms existing methods.

Text-to-video diffusion models enable the generation of high-quality videos that follow text instructions, simplifying the process of producing diverse and individual content. Current methods excel in generating short videos (up to 16s), but produce hard-cuts when naively extended to long video synthesis. To overcome these limitations, we present StreamingT2V, an autoregressive method that generates long videos of up to 2 minutes or longer with seamless transitions. The key components are: (i) a short-term memory block called conditional attention module (CAM), which conditions the current generation on the features extracted from the preceding chunk via an attentional mechanism, leading to consistent chunk transitions, (ii) a long-term memory block called appearance preservation module (APM), which extracts high-level scene and object features from the first video chunk to prevent the model from forgetting the initial scene, and (iii) a randomized blending approach that allows for the autoregressive application of a video enhancer on videos of indefinite length, ensuring consistency across chunks. Experiments show that StreamingT2V produces more motion, while competing methods suffer from video stagnation when applied naively in an autoregressive fashion. Thus, we propose with StreamingT2V a high-quality seamless text-to-long video generator, surpassing competitors in both consistency and motion.

Recent generative models have demonstrated impressive capabilities in generating realistic and visually pleasing images grounded on textual prompts. Nevertheless, a significant challenge remains in applying these models for the more intricate task of story visualization. Since it requires resolving pronouns (he, she, they) in the frame descriptions, i.e., anaphora resolution, and ensuring consistent characters and background synthesis across frames. Yet, the emerging Large Language Model (LLM) showcases robust reasoning abilities to navigate through ambiguous references and process extensive sequences. Therefore, we introduce StoryGPT-V, which leverages the merits of the latent diffusion (LDM) and LLM to produce images with consistent and high-quality characters grounded on given story descriptions. First, we train a character-aware LDM, which takes character-augmented semantic embedding as input and includes the supervision of the cross-attention map using character segmentation masks, aiming to enhance character generation accuracy and faithfulness. In the second stage, we enable an alignment between the output of LLM and the character-augmented embedding residing in the input space of the first-stage model. This harnesses the reasoning ability of LLM to address ambiguous references and the comprehension capability to memorize the context. We conduct comprehensive experiments on two visual story visualization benchmarks. Our model reports superior quantitative results and consistently generates accurate characters of remarkable quality with low memory consumption. Our code is publicly available at: https://xiaoqian-shen.github.io/StoryGPT-V.

Projected Gradient Descent (PGD) under the L_infinity ball has become one of the defacto methods used in adversarial robustness evaluation for computer vision (CV) due to its reliability and efficacy, making a strong and easy-to-implement iterative baseline. However, PGD is computationally demanding to apply, especially when thousands of iterations is current best-practice recommendations to generate an adversarial example for a single image. In this work, we introduce a simple novel method for early termination of PGD based on cycle detection by exploiting the geometry of how PGD is implemented in practice and show that it can produce large speedup factors while providing the exact same estimate of model robustness as standard PGD. This method substantially speeds up PGD without sacrificing any attack strength, enabling evaluations of robustness that were previously computationally intractable

Multimodal Large Language Models (MLLMs) frequently suffer from hallucination issues, generating information about objects that are not present in input images during vision-language tasks. These hallucinations particularly undermine model reliability in practical applications requiring accurate object identification. To address this challenge, we propose TL-DPO, a preference learning approach that mitigates hallucinations by focusing on targeted areas where they occur. To implement this, we build a dataset containing hallucinated responses, correct responses, and target information (i.e., objects present in the images and the corresponding chunk positions in responses affected by hallucinations). By applying a preference learning method restricted to these specific targets, the model can filter out irrelevant signals and focus on correcting hallucinations. This allows the model to produce more factual responses by concentrating solely on relevant information. Experimental results demonstrate that TL-DPO effectively reduces hallucinations across multiple vision hallucination tasks, improving the reliability and performance of MLLMs without diminishing overall performance.

Action-driven stochastic human motion prediction aims to generate future motion sequences of a pre-defined target action based on given past observed sequences performing non-target actions. This task primarily presents two challenges. Firstly, generating smooth transition motions is hard due to the varying transition speeds of different actions. Secondly, the action characteristic is difficult to be learned because of the similarity of some actions. These issues cause the predicted results to be unreasonable and inconsistent. As a result, we propose two memory banks, the Soft-transition Action Bank (STAB) and Action Characteristic Bank (ACB), to tackle the problems above. The STAB stores the action transition information. It is equipped with the novel soft searching approach, which encourages the model to focus on multiple possible action categories of observed motions. The ACB records action characteristic, which produces more prior information for predicting certain actions. To fuse the features retrieved from the two banks better, we further propose the Adaptive Attention Adjustment (AAA) strategy. Extensive experiments on four motion prediction datasets demonstrate that our approach consistently outperforms the previous state-of-the-art. The demo and code are available at https://hyqlat.github.io/STABACB.github.io/.