CVPR 2025 Papers

The latest emerged 4D Panoptic Scene Graph (4D-PSG) provides an advanced-ever representation for comprehensively modeling the dynamic 4D visual real world. Unfortunately, current pioneering 4D-PSG research can largely suffer from data scarcity issues severely, as well as the resulting out-of-vocabulary problems; also, the pipeline nature of the benchmark generation method can lead to suboptimal performance. To address these challenges, this paper investigates a novel framework for 4D-PSG generation that leverages rich 2D visual scene annotations to enhance 4D scene learning. First, we introduce a 4D Large Language Model (4D-LLM) integrated with a 3D mask decoder for end-to-end generation of 4D-PSG. A chained SG inference mechanism is further designed to exploit LLMs' open-vocabulary capabilities to infer accurate and comprehensive object and relation labels iteratively. Most importantly, we propose a 2D-to-4D visual scene transfer learning framework, where a spatial-temporal scene transcending strategy effectively transfers dimension-invariant features from abundant 2D SG annotations to 4D scenes, effectively compensating for data scarcity in 4D-PSG. Extensive experiments on the benchmark data demonstrate that we strikingly outperform baseline models by a large margin, highlighting the effectiveness of our method.

Learned image compression methods have attracted great research interest and exhibited superior rate-distortion performance to the best classical image compression standards of the present.The entropy model plays a key role in learned image compression, which estimates the probability distribution of the latent representation for further entropy coding.Most existing methods employed hyper-prior and auto-regressive architectures to form their entropy models.However, they only aimed to explore the internal dependencies of latent representation while neglecting the importance of extracting prior from training data.In this work, we propose a novel entropy model named Dictionary-based Cross Attention Entropy model, which introduces a learnable dictionary to summarize the typical structures occurring in the training dataset to enhance the entropy model.Extensive experimental results have demonstrated that the proposed model strikes a better balance between performance and latency, achieving state-of-the-art results on various benchmark datasets.

Extracting high-fidelity RGBD information from two-dimensional (2D) images is essential for various visual computing applications. Stereo imaging, as a reliable passive imaging technique for obtaining three-dimensional (3D) scene information, has benefited greatly from deep learning advancements. However, existing stereo depth estimation algorithms struggle to perceive high-frequency information and resolve high-resolution depth maps in realistic camera settings with large depth variations. These algorithms commonly neglect the hardware parameter configuration, limiting the potential for achieving optimal solutions solely through software-based design strategies.This work presents a hardware-software co-designed RGBD imaging framework that leverages both stereo and focus cues to reconstruct texture-rich color images along with detailed depth maps over a wide depth range. A pair of rank-2 parameterized diffractive optical elements (DOEs) is employed to encode perpendicular complementary information optically during stereo acquisitions. Additionally, we employ an IGEV-UNet-fused neural network tailored to the proposed rank-2 encoding for stereo matching and image reconstruction. Through prototyping a stereo camera with customized DOEs, our deep stereo imaging paradigm has demonstrated superior performance over existing monocular and stereo imaging systems in both image PSNR by 2.96 dB gain and depth accuracy in high-frequency details across distances from 0.67 to 8 meters.

Capturing the temporal evolution of Gaussian properties such as position, rotation, and scale is a challenging task due to the vast number of time-varying parameters and the limited photometric data available, which generally results in convergence issues, making it difficult to find an optimal solution. While feeding all inputs into an end-to-end neural network can effectively model complex temporal dynamics, this approach lacks explicit supervision and struggles to generate high-quality transformation fields. On the other hand, using time-conditioned polynomial functions to model Gaussian trajectories and orientations provides a more explicit and interpretable solution, but requires significant handcrafted effort and lacks generalizability across diverse scenes. To overcome these limitations, this paper introduces a novel approach based on a learnable infinite Taylor Formula to model the temporal evolution of Gaussians. This method offers both the flexibility of an implicit network-based approach and the interpretability of explicit polynomial functions, allowing for more robust and generalizable modeling of Gaussian dynamics across various dynamic scenes.Extensive experiments on dynamic novel view rendering tasks are conducted on public datasets, demonstrating that the proposed method achieves state-of-the-art performance in this domain.

Rencently, Gaussian splatting has demonstrated significant success in novel view synthesis. Current methods often regress Gaussians with pixel or point cloud correspondence, linking each Gaussian with a pixel or a 3D point. This leads to the redundancy of Gaussians being used to overfit the correspondence rather than the objects represented by the 3D Gaussians themselves, consequently wasting resources and lacking accurate geometries or textures.In this paper, we introduce LeanGaussian, a novel approach that treats each query in deformable Transformer as one 3D Gaussian ellipsoid, breaking the pixel or point cloud correspondence constraints. We leverage deformable decoder to iteratively refine the Gaussians layer-by-layer with the image features as keys and values.Notably, the center of each 3D Gaussian is defined as 3D reference points, which are then projected onto the image for deformable attention in 2D space.On both the ShapeNet SRN dataset (category level) and the Google Scanned Objects dataset (open-category level, trained with the Objaverse dataset), our approach, outperforms prior methods by approximately 6.1%, achieving a PSNR of 25.44 and 22.36, respectively. Additionally, our method achieves a 3D reconstruction speed of 7.2 FPS and rendering speed 500 FPS.

Spatial reasoning is a fundamental aspect of human cognition, enabling intuitive understanding and manipulation of objects in three-dimensional space. While foundation models demonstrate remarkable performance on some benchmarks, they still struggle with 3D reasoning tasks like arranging objects in space according to open-ended language instructions, particularly in dense and physically constrained environments. We introduce LayoutVLM, a framework and scene layout representation that exploits the semantic knowledge of Vision-Language Models (VLMs) and supports differentiable optimization to ensure physical plausibility. LayoutVLM employs VLMs to generate two mutually reinforcing representations from visually marked images, and a self-consistent decoding process to improve VLMs spatial planning. Our experiments show that LayoutVLM addresses the limitations of existing LLM and constraint-based approaches, producing physically plausible 3D layouts better aligned with the semantic intent of input language instructions. We also demonstrate that fine-tuning VLMs with the proposed scene layout representation extracted from existing scene datasets can improve their reasoning performance.

Computer vision is largely based on 2D techniques, with 3D vision still relegated to a relatively narrow subset of applications. However, by building on recent advances in 3D models such as neural radiance fields, some authors have shown that 3D techniques can at last improve outputs extracted from independent 2D views, by fusing them into 3D and denoising them. This is particularly helpful in egocentric videos, where the camera motion is significant, but only under the assumption that the scene itself is static. In fact, as shown in the recent analysis conducted by EPIC Fields, 3D techniques are ineffective when it comes to studying dynamic phenomena, and, in particular, when segmenting moving objects. In this paper, we look into this issue in more detail. First, we propose to improve dynamic segmentation in 3D by fusing motion segmentation predictions from a 2D-based model into layered radiance fields (Layered Motion Fusion). However, the high complexity of long, dynamic videos makes it challenging to capture the underlying geometric structure, and, as a result, hinders the fusion of motion cues into the (incomplete) scene geometry. We address this issue through test-time refinement, which helps the model to focus on specific frames, thereby reducing the data complexity. This results in a synergy between motion fusion and the refinement, and in turn leads to segmentation predictions of the 3D model that surpass the 2D baseline by a large margin. This demonstrates that 3D techniques can enhance 2D analysis even for dynamic phenomena in a challenging and realistic setting.

This work presents a progressive image vectorization technique that reconstructs the raster image as layer-wise vectors from semantic-aligned macro structures to finer details. Our approach introduces a new image simplification method leveraging the feature-average effect in the Score Distillation Sampling mechanism, achieving effective visual abstraction from the detailed to coarse. Guided by the sequence of progressive simplified images, we propose a two-stage vectorization process of structural buildup and visual refinement, constructing the vectors in an organized and manageable manner. The resulting vectors are layered and well-aligned with the target image's explicit and implicit semantic structures. Our method demonstrates high performance across a wide range of images. Comparative analysis with existing vectorization methods highlights our technique's superiority in creating vectors with high visual fidelity, and more importantly, achieving higher semantic alignment and more compact layered representation.

Diffusion Transformers (DiTs) have achieved state-of-the-art (SOTA) image generation quality but suffer from high latency and memory inefficiency, making them difficult to deploy on resource-constrained devices. One major efficiency bottleneck is that existing DiTs apply equal computation across all regions of an image. However, not all image tokens are equally important, and certain localized areas require more computation, such as objects. To address this, we propose DiffCR, a dynamic DiT inference framework with differentiable compression ratios, which automatically learns to dynamically route computation across layers and timesteps for each image token, resulting in efficient DiTs. Specifically, DiffCR integrates three features: (1) A token-level routing scheme where each DiT layer includes a router that is fine-tuned jointly with model weights to predict token importance scores. In this way, unimportant tokens bypass the entire layer's computation; (2) A layer-wise differentiable ratio mechanism where different DiT layers automatically learn varying compression ratios from a zero initialization, resulting in large compression ratios in redundant layers while others remain less compressed or even uncompressed; (3) A timestep-wise differentiable ratio mechanism where each denoising timestep learns its own compression ratio. The resulting pattern shows higher ratios for noisier timesteps and lower ratios as the image becomes clearer. Extensive experiments on text-to-image and inpainting tasks show that DiffCR effectively captures dynamism across token, layer, and timestep axes, achieving superior trade-offs between generation quality and efficiency compared to prior works. The project website is available at https://www.haoranyou.com/diffcr.

Current research on generating 3D hand-object interaction motion primarily focuses on in-domain objects. Generalization to unseen objects is essential for practical applications, yet it remains both challenging and largely unexplored.In this paper, we propose LatentHOI, a novel approach designed to tackle the challenges of generalizing hand-object interaction to unseen objects.Our main insight lies in decoupling high-level temporal motion from fine-grained spatial hand-object interactions with a latent diffusion model coupled with a Grasping Variational Autoencoder (GraspVAE). This configuration not only enhances the conditional dependency between spatial grasp and temporal motion but also improves data utilization and reduces overfitting through regularization in the latent space. We conducted extensive experiments in an unseen-object setting on both single-hand grasping and bi-manual motion datasets, including GRAB, DexYCB, and OakInk.Quantitative and qualitative evaluations demonstrate that our method significantly enhances the realism and physical plausibility of generated motions for unseen objects, both in single and bimanual manipulations, compared to the state-of-the-art.

In this paper, we propose LSRNA, a novel framework for higher-resolution (exceeding 1K) image generation using diffusion models by leveraging super-resolution directly in the latent space. Existing diffusion models struggle with scaling beyond their training resolutions, often leading to structural distortions or content repetition. Reference-based methods address the issues by upsampling a low-resolution reference to guide higher-resolution generation. However, they face significant challenges: upsampling in latent space often causes manifold deviation, which degrades output quality. On the other hand, upsampling in RGB space tends to produce overly smoothed outputs. To overcome these limitations, LSRNA combines Latent space Super-Resolution (LSR) for manifold alignment and Region-wise Noise Addition (RNA) to enhance high-frequency details. Our extensive experiments demonstrate that integrating LSRNA outperforms state-of-the-art reference-based methods across various resolutions and metrics, while showing the critical role of latent space upsampling in preserving detail and sharpness. The code will be available at https://github.com/3587jjh/LSRNA https://github.com/3587jjh/LSRNA.

Digital imaging systems have traditionally relied on brute-force measurement and processing of pixels arranged on regular grids. In contrast, the human visual system performs significant data reduction from the large number of photoreceptors to the optic nerve, effectively encoding visual information into a low-bandwidth latent space representation optimized for brain processing. Inspired by this, we propose a similar approach to advance artificial vision systems. Latent Space Imaging introduces a new paradigm that combines optics and software to encode image information directly into the semantically rich latent space of a generative model. This approach substantially reduces bandwidth and memory demands during image capture and enables a range of downstream tasks focused on the latent space.We validate this principle through an initial hardware prototype based on a single-pixel camera. By implementing an amplitude modulation scheme that encodes into the generative model's latent space, we achieve compression ratios ranging from 1:100 to 1:1000 during imaging, and up to 1:16384 for downstream applications. This approach leverages the model's intrinsic linear boundaries, demonstrating the potential of latent space imaging for highly efficient imaging hardware, adaptable future applications in high-speed imaging, and task-specific cameras with significantly reduced hardware complexity.

Scaling by training on large datasets has been shown to enhance the quality and fidelity of image generation and manipulation with diffusion models; however, such large datasets are not always accessible in medical imaging due to cost and privacy issues, which contradicts one of the main applications of such models to produce synthetic samples where real data is scarce. Also, finetuning on pre-trained general models has been a challenge due to the distribution shift between the medical domain and the pre-trained models. Here, we propose Latent Drift (LD) for diffusion models that can be adopted for any fine-tuning method to mitigate the issues faced by the distribution shift or employed in inference time as a condition. Latent Drifting enables diffusion models to be conditioned for medical images fitted for the complex task of counterfactual image generation, which is crucial to investigate how parameters such as gender, age, and adding or removing diseases in a patient would alter the medical images. We evaluate our method on three public longitudinal benchmark datasets of brain MRI and chest X-rays for counterfactual image generation. Our results demonstrate significant performance gains in various scenarios when combined with different fine-tuning schemes. The source code of this work is available on GitHub.

Multi-view clustering is a long-standing hot topic in machine learning communities, due to its capability of integrating data information from multiple sources and modalities. By utilizing tensor Singular Value Decomposition (t-SVD) technique with the tensor rotation trick, recent advances have achieved remarkable improvements on clustering performance. However, we find this is attributed to the inadvertent use of sequential information of sorted data samples, i.e. inadvertent label use, which violates the unsupervised learning setting. On the other hand, existing large-scale approaches are mostly developed on the basis of matrix factorization or anchor techniques, thereby fail to consider the similarities among all data samples, preventing from further performance improvement. To address the above issues, we first analyze the tensor rotation trick and recommend to remove it from tensor clustering. On its basis, a novel large-scale multi-view tensor clustering method is developed by incorporating the pair-wise similarities with implicit linear kernel function. To solve the resultant optimization problem, we design an efficient algorithm of linear complexity.Moreover, extensive experiments are conducted and corresponding results well support the aforementioned finding and validate the effectiveness and efficiency of the proposed method.

Subject-driven text-to-image generation aims to produce images of a new subject within a desired context by accurately capturing both the visual characteristics of the subject and the semantic content of a text prompt. Traditional methods rely on time- and resource-intensive fine-tuning for subject alignment, while recent zero-shot approaches leverage on-the-fly image prompting, often sacrificing subject alignment. In this paper, we introduce Diptych Prompting, a novel zero-shot approach that reinterprets as an inpainting task with precise subject alignment by leveraging the emergent property of diptych generation in large-scale text-to-image models. Diptych Prompting arranges an incomplete diptych with the reference image in the left panel, and performs text-conditioned inpainting on the right panel. We further prevent unwanted content leakage by removing the background in the reference image and improve fine-grained details in the generated subject by enhancing attention weights between the panels during inpainting. Experimental results confirm that our approach significantly outperforms zero-shot image prompting methods, resulting in images that are visually preferred by users. Additionally, our method supports not only subject-driven generation but also stylized image generation and subject-driven image editing, demonstrating versatility across diverse image generation applications.

The current state-of-the-art methods in domain adaptive object detection (DAOD) use Mean Teacher self-labelling, where a teacher model, directly derived as an exponential moving average of the student model, is used to generate labels on the target domain which are then used to improve both models in a positive loop. This couples learning and generating labels on the target domain, and other recent works also leverage the generated labels to add additional domain alignment losses. We believe this coupling is brittle and excessively constrained: there is no guarantee that a student trained only on source data can generate accurate target domain labels and initiate the positive feedback loop, and much better target domain labels can likely be generated by using a large pretrained network that has been exposed to much more data. Vision foundational models are exactly such models, and they have shown impressive task generalization capabilities even when frozen. We want to leverage these models for DAOD and introduce DINO Teacher, which consists of two components. First, we train a new labeller on source data only using a large frozen DINOv2 backbone and show it generates more accurate labels than Mean Teacher. Next, we align the student's source and target image patch features with those from a DINO encoder, driving source and target representations closer to the generalizable DINO representation. We obtain state-of-the-art performance on multiple DAOD datasets.

Salient Object Ranking (SOR) aims to study human attention shifts across different objects in the scene. It is a challenging task, as it requires comprehension of the relations among the salient objects in the scene. However, existing works often overlook such relations or model them implicitly. In this work, we observe that when Large Vision-Language Models (LVLMs) describe a scene, they usually focus on the most salient object first, and then discuss the relations as they move on to the next (less salient) one. Based on this observation, we propose a novel Language-Guided Salient Object Ranking approach (named LG-SOR), which utilizes the internal knowledge within the LVLM-generated language descriptions, i.e., semantic relation cues and the implicit entity order cues, to facilitate saliency ranking. Specifically, we first propose a novel Text-Guided Visual Modulation (TGVM) module to incorporate semantic information in the description for saliency ranking. TGVM controls the flow of linguistic information to the visual features, suppresses noisy background image features, and enables propagation of useful textual features. We then propose a novel Text-Aware Visual Reasoning (TAVR) module to enhance model reasoning in object ranking, by explicitly learning a multimodal graph based on the entity and relation cues derived from the description. Extensive experiments demonstrate superior performances of our model on two SOR benchmarks.

Image tokenization, the process of transforming raw image pixels into a compact low-dimensional latent representation, has proven crucial for scalable and efficient image generation. However, mainstream image tokenization methods generally have limited compression rates, making high-resolution image generation computationally expensive. To address this challenge, we propose to leverage language for efficient image tokenization, and we call our method Text-Conditioned Image Tokenization (TexTok). TexTok is a simple yet effective tokenization framework that leverages language to provide a compact, high-level semantic representation. By conditioning the tokenization process on descriptive text captions, TexTok simplifies semantic learning, allowing more learning capacity and token space to be allocated to capture fine-grained visual details, leading to enhanced reconstruction quality and higher compression rates. Compared to the conventional tokenizer without text conditioning, TexTok achieves average reconstruction FID improvements of 29.2% and 48.1% on ImageNet-256 and -512 benchmarks respectively, across varying numbers of tokens. These tokenization improvements consistently translate to 16.3% and 34.3% average improvements in generation FID. By simply replacing the tokenizer in Diffusion Transformer (DiT) with TexTok, our system can achieve a 93.5x inference speedup while still outperforming the original DiT using only 32 tokens on ImageNet-512. TexTok with a vanilla DiT generator achieves state-of-the-art FID scores of 1.46 and 1.62 on ImageNet-256 and -512 respectively. Furthermore, we demonstrate TexTok's superiority on the text-to-image generation task, effectively utilizing the off-the-shelf text captions in tokenization.

Long-term sports assessment is a challenging task in video understanding since it requires judging complex movement variations and action-music coordination. However, there is no direct correlation between the diverse background music and movements in sporting events. Previous works require a large number of model parameters to learn potential associations between actions and music. To address this issue, we propose a language-guided audio-visual learning (MLAVL) framework that models "audio-action-visual" correlations guided by low-cost language modality. In our framework, multidimensional domain-based actions form action knowledge graphs, motivating audio-visual modalities to focus on task-relevant actions. We further design a shared-specific context encoder to integrate deep multimodal semantics, and an audio-visual cross-modal fusion module to evaluate action-music consistency. To match the sport's rules, we then propose a dual-branch prompt-guided grading module to weigh both visual and audio-visual performance. Extensive experiments demonstrate that our approach achieves state-of-the-art on four public long-term sports benchmarks while maintaining low parameters. Our code is available at https://github.com/XuHuangbiao/MLAVL.