CVPR 2025 Papers

Diffusion models have achieved remarkable success in text-to-image generation. However, their practical applications are hindered by the misalignment between generated images and corresponding text prompts. To tackle this issue, reinforcement learning (RL) has been considered for diffusion model fine-tuning. Yet, RL's effectiveness is limited by the challenge of sparse reward, where feedback is only available at the end of the generation process. This makes it difficult to identify which actions during the denoising process contribute positively to the final generated image, potentially leading to ineffective or unnecessary denoising policies. To this end, this paper presents a novel RL-based framework that addresses the sparse reward problem when training diffusion models. Our framework, named \text B ^2\text -DiffuRL , employs two strategies: **B**ackward progressive training and **B**ranch-based sampling. For one thing, backward progressive training focuses initially on the final timesteps of the denoising process and gradually extends the training interval to earlier timesteps, easing the learning difficulty associated with sparse rewards. For another, we perform branch-based sampling for each training interval. By comparing the samples within the same branch, we can identify how much the policies of the current training interval contribute to the final image, which helps to learn effective policies instead of unnecessary ones. \text B ^2\text -DiffuRL is compatible with existing optimization algorithms. Extensive experiments demonstrate the effectiveness of \text B ^2\text -DiffuRL in improving prompt-image alignment and maintaining diversity in generated images. The code for this work is available.

Micromobility, which utilizes lightweight devices moving in urban public spaces - such as delivery robots and electric wheelchairs - emerges as a promising alternative to vehicular mobility. Current micromobility depends mostly on human manual operation (in-person or remote control), which raises safety and efficiency concerns when navigating busy urban environments full of obstacles and pedestrians. Assisting humans with AI agents in maneuvering micromobility devices presents a viable solution for enhancing safety and efficiency. In this work, we present a scalable urban simulation solution to advance autonomous micromobility. First, we build URBAN-SIM -- a high-performance robot learning platform for large-scale training of embodied agents in interactive urban scenes. URBAN-SIM contains three critical modules: Hierarchical Urban Generation pipeline, Interactive Dynamics Generation strategy, and Asynchronous Scene Sampling scheme, to improve the diversity, realism, and efficiency of robot learning in simulation. Then, we propose URBAN-BENCH -- a suite of essential tasks and benchmarks to gauge various capabilities of the AI agents in achieving autonomous micromobility. URBAN-BENCH includes eight tasks based on three core skills of the agents: Urban Locomotion, Urban Navigation, and Urban Traverse. We evaluate four robots with heterogeneous embodiments, such as the wheeled and legged robots, across these tasks. Experiments on diverse terrains and urban structures reveal each robot's unique strengths and limitations. This work will be open-sourced and under sustainable maintenance to foster future research in autonomous micromobility.

Medical image re-identification (MedReID) is under-explored so far, despite its critical applications in personalized healthcare and privacy protection.In this paper, we introduce a thorough benchmark and a unified model for this problem.First, to handle various medical modalities, we propose a novel Continuous Modality-based Parameter Adapter (ComPA). ComPA condenses medical content into a continuous modality representation and dynamically adjusts the modality-agnostic model with modality-specific parameters at runtime. This allows a single model to adaptively learn and process diverse modality data.Furthermore, we integrate medical priors into our model by aligning it with a bag of pre-trained medical foundation models, in terms of the differential features.Compared to single-image feature, modeling the inter-image difference better fits the re-identification problem, which involves discriminating multiple images.We evaluate the proposed model against 25 foundation models and 8 large multi-modal language models across 11 image datasets, demonstrating consistently superior performance.Additionally, we deploy the proposed MedReID technique to two real-world applications, i.e., history-augmented personalized diagnosis and medical privacy protection.

We consider the challenging problem of learning Signed Distance Functions (SDF) from sparse and noisy 3D point clouds. In contrast to recent methods that depend on smoothness priors, our method, rooted in a distributionally robust optimization (DRO) framework, incorporates a regularization term that leverages samples from the uncertainty regions of the model to improve the learned SDFs. Thanks to tractable dual formulations, we show that this framework enables a stable and efficient optimization of SDFs in the absence of ground truth supervision. Using a variety of synthetic and real data evaluations from different modalities, we show that of our DRO based learning framework can improve SDF learning with respect to baselines and the state-of-the-art.

Bird's-eye view (BEV) perception has gained significant attention because it provides a unified representation to fuse multiple view images and enables a wide range of downstream autonomous driving tasks, such as forecasting and planning. Recent state-of-the-art models utilize projection-based methods which formulate BEV perception as query learning to bypass explicit depth estimation. While we observe promising advancements in this paradigm, they still fall short of real-world applications because of the lack of uncertainty modeling and expensive computational requirement. In this work, we introduce GaussianLSS, an uncertainty-aware BEV perception framework that revisits the unprojection-based method, specifically the Lift-Splat-Shoot (LSS) paradigm, and enhances it with depth uncertainty modeling. Our GaussianLSS represents spatial dispersion by learning a soft depth mean and computing the variance of the depth distribution, which implicitly captures object extents. We then transform the depth distribution into 3D Gaussians and rasterize them to construct uncertainty-aware BEV features. We evaluate GaussianLSS on the nuScenes dataset, achieving state-of-the-art performance compared to unprojection-based methods. In particular, it provides significant advantages in speed, running 2x faster, and in memory efficiency, using 0.3x less memory compared to projection-based methods, while achieving competitive performance with only a 0.7% IoU difference.

Full-reference image quality assessment (FR-IQA) generally assumes that reference images are of perfect quality. However, this assumption is flawed due to the sensor and optical limitations of modern imaging systems. Moreover, recent generative enhancement methods are capable of producing images of higher quality than their original. All of these challenge the effectiveness and applicability of current FR-IQA models. To relax the assumption of perfect reference image quality, we build a large-scale IQA database, namely DiffIQA, containing approximately 180,000 images generated by a diffusion-based image enhancer with adjustable hyper-parameters. Each image is annotated by human subjects as either worse, similar, or better quality compared to its reference. Building on this, we present a generalized FR-IQA model, namely Adaptive FIdelity-Naturalness Evaluator (A-FINE), to accurately assess and adaptively combine the fidelity and naturalness of a test image. A-FINE aligns well with standard FR-IQA when the reference image is much more natural than the test image. We demonstrate by extensive experiments that A-FINE surpasses standard FR-IQA models on well-established IQA datasets and our newly created DiffIQA. To further validate A-FINE, we additionally construct a super-resolution IQA benchmark (SRIQA-Bench), encompassing test images derived from ten state-of-the-art SR methods with reliable human quality annotations. Tests on SRIQA-Bench re-affirm the advantages of A-FINE. The code and dataset are available at https://tianhewu.github.io/A-FINE-page.github.io/.

Diffusion models have made breakthroughs in 3D generation tasks. Current 3D diffusion models focus on reconstructing target shape from images or a set of partial observations. While excelling in global context understanding, they struggle to capture the local details of complex shapes and limited to the occlusion and lighting conditions. To overcome these limitations, we utilize tactile images to capture the local 3D information and propose a Touch2Shape model, which leverages a touch-conditioned diffusion model to explore and reconstruct the target shape from touch. For shape reconstruction, we have developed a touch embedding module to condition the diffusion model in creating a compact representation and a touch shape fusion module to refine the reconstructed shape. For shape exploration, we combine the diffusion model with reinforcement learning to train a policy. This involves using the generated latent vector from the diffusion model to guide the touch exploration policy training through a novel reward design. Experiments validate the reconstruction quality thorough both qualitatively and quantitative analysis, and our touch exploration policy further boosts reconstruction performance.

Recent advancements in Diffusion Transformer (DiT) have demonstrated remarkable proficiency in producing high-quality video content. Nonetheless, the potential of transformer-based diffusion models for effectively generating videos with controllable motion remains an area of limited exploration. This paper introduces Tora, the first trajectory-oriented DiT framework that concurrently integrates textual, visual, and trajectory conditions, thereby enabling scalable video generation with effective motion guidance. Specifically, Tora consists of a Trajectory Extractor (TE), a Spatial-Temporal DiT, and a Motion-guidance Fuser (MGF). The TE encodes arbitrary trajectories into hierarchical spacetime motion patches with a 3D motion compression network. The MGF integrates the motion patches into the DiT blocks to generate consistent videos that accurately follow designated trajectories. Our design aligns seamlessly with DiT's scalability, allowing precise control of video content's dynamics with diverse durations, aspect ratios, and resolutions. Extensive experiments demonstrate that Tora excels in achieving high motion fidelity compared to the foundational DiT model, while also accurately simulating the complex movements of the physical world. Code is made available at https://github.com/alibaba/Tora .

Accurately modeling multi-class cell topology is crucial in digital pathology, as it provides critical insights into tissue structure and pathology. The synthetic generation of cell topology enables realistic simulations of complex tissue environments, enhances downstream tasks by augmenting training data, aligns more closely with pathologists' domain knowledge, and offers new opportunities for controlling and generalizing the tumor microenvironment. In this paper, we propose a novel approach that integrates topological constraints into a diffusion model to improve the generation of realistic, contextually accurate cell topologies. Our method refines the simulation of cell distributions and interactions, increasing the precision and interpretability of results in downstream tasks such as cell detection and classification. To assess the topological fidelity of generated layouts, we introduce a new metric, Topological Frechet Distance (TopoFD), which overcomes the limitations of traditional metrics like FID in evaluating topological structure. Experimental results demonstrate the effectiveness of our approach in generating multi-class cell layouts that capture intricate topological relationships. Code is available at https://github.com/Melon-Xu/TopoCellGen.

Vision-Language Models (VLMs) demand substantial computational resources during inference, largely due to the extensive visual input tokens for representing visual information. Previous studies have noted that visual tokens tend to receive less attention than text tokens, suggesting their lower importance during inference and potential for pruning. However, their methods encounter several chal- lenges: reliance on greedy heuristic criteria for token im- portance and incompatibility with FlashAttention and KV cache. To address these issues, we introduce TopV, a com- patible TOken Pruning with inference Time Optimization for fast and low-memory VLM, achieving efficient pruning without additional training or fine-tuning. Instead of re- lying on attention scores, we formulate token pruning as an optimization problem, accurately identifying important visual tokens while remaining compatible with FlashAtten- tion. Additionally, since we only perform this pruning once during the prefilling stage, it effectively reduces KV cache size. Our optimization framework incorporates a visual- aware cost function considering factors such as Feature Similarity, Relative Spatial Distance, and Absolute Central Distance, to measure the importance of each source visual token, enabling effective pruning of low-importance tokens. Extensive experiments demonstrate that our method outperforms previous token pruning methods, validating the effectiveness and efficiency of our approach.

Efficient Point Cloud Geometry Compression (PCGC) with a lower bits per point (BPP) and higher peak signal-to-noise ratio (PSNR) is essential for the transportation of large-scale 3D data. Although octree-based entropy models can reduce BPP without introducing geometry distortion, existing CNN-based models struggle with limited receptive fields to capture long-range dependencies, while Transformer-built architectures always neglect fine-grained details due to their reliance on global self-attention. In this paper, we propose a Transformer-efficient occupancy prediction Network, termed TopNet, to overcome these challenges by developing several novel components: Locally-enhanced Context Encoding (LeCE) for enhancing the translation-invariance of the octree nodes, Adaptive-Length Sliding Window Attention (AL-SWA) for capturing both global and local dependencies while adaptively adjusting attention weights based on the input window length, Spatial-Gated-enhanced Channel Mixer (SG-CM) for efficient feature aggregation from ancestors and siblings, and Latent-guided Node Occupancy Predictor (LNOP) for improving prediction accuracy of spatially adjacent octree nodes. Comprehensive experiments across both indoor and outdoor point cloud datasets demonstrate that our TopNet achieves state-of-the-art performance with fewer parameters, further advancing the reduction-efficiency boundaries of PCGC. The code is available at https://github.com/xinjiewang1995/TopNet.

Global contextual information and local detail features are essential for haze removal tasks. Deep learning models perform well on small, low-resolution images, but they encounter difficulties with large, high-resolution ones due to GPU memory limitations. As a compromise, they often resort to image slicing or downsampling. The former diminishes global information, while the latter discards high-frequency details. To address these challenges, we propose DehazeXL, a haze removal method that effectively balances global context and local feature extraction, enabling end-to-end modeling of large images on mainstream GPU hardware. Additionally, to evaluate the efficiency of global context utilization in haze removal performance, we design a visual attribution method tailored to the characteristics of haze removal tasks. Finally, recognizing the lack of benchmark datasets for haze removal in large images, we have developed an ultra-high-resolution haze removal dataset (8KDehaze) to support model training and testing. It includes 10000 pairs of clear and hazy remote sensing images, each sized at 8192 x8192 pixels. Extensive experiments demonstrate that DehazeXL can infer images up to 10240 x10240 pixels with only 21 GB of memory, achieving state-of-the-art results among all evaluated methods. The source code and experimental dataset will soon be made publicly available.

Human-centric motion control in video generation remains a critical challenge, particularly when jointly controlling camera movements and human poses in scenarios like the iconic Grammy Glambot moment. While recent video diffusion models have made significant progress, existing approaches struggle with limited motion representations and inadequate integration of camera and human motion controls. In this work, we present TokenMotion, the first DiT-based video diffusion framework that enables fine-grained control over camera motion, human motion, and their joint interaction. We represent camera trajectories and human poses as spatio-temporal tokens to enable local control granularity. Our approach introduces a unified modeling framework utilizing a decouple-and-fuse strategy, bridged by a human-aware dynamic mask that effectively handles the spatially varying nature of combined motion signals. Through extensive experiments, we demonstrate TokenMotion's effectiveness across both text-to-video and image-to-video paradigms, consistently outperforming current state-of-the-art methods in human-centric motion control tasks. Our work represents a significant advancement in controllable video generation, with particular relevance for creative production applications ranging from films to short-form content.

Synthesizing diverse and physically plausible Human-Scene Interactions (HSI) is pivotal for both computer animation and embodied AI. Despite encouraging progress, current methods mainly focus on developing separate controllers, each specialized for a specific interaction task. This significantly hinders the ability to tackle a wide variety of challenging HSI tasks that require the integration of multiple skills, e.g., sitting down while carrying an object (see Fig. 1). To address this issue, we present TokenHSI, a single, unified transformer-based policy capable of multi-skill unification and flexible adaptation. The key insight is to model the humanoid proprioception as a separate shared token and combine it with distinct task tokens via a masking mechanism. Such a unified policy enables effective knowledge sharing across skills, thereby facilitating the multi-task training. Moreover, our policy architecture supports variable length inputs, enabling flexible adaptation of learned skills to new scenarios. By training additional task tokenizers, we can not only modify the geometries of interaction targets but also coordinate multiple skills to address complex tasks. The experiments demonstrate that our approach can significantly improve versatility, adaptability, and extensibility in various HSI tasks.

We present TokenFlow, a novel unified image tokenizer that bridges the long-standing gap between multimodal understanding and generation. Prior research attempt to employ a single reconstruction-targeted Vector Quantization (VQ) encoder for unifying these two tasks. We observe that understanding and generation require fundamentally different granularities of visual information. This leads to a critical trade-off, particularly compromising performance in multimodal understanding tasks. TokenFlow addresses this challenge through an innovative dual-codebook architecture that decouples semantic and pixel-level feature learning while maintaining their alignment via a shared mapping mechanism. This design enables direct access to both high-level semantic representations crucial for understanding tasks and fine-grained visual features essential for generation through shared indices. Our extensive experiments demonstrate TokenFlow's superiority across multiple dimensions. Leveraging TokenFlow, we demonstrate for the first time that discrete visual input can surpass LLaVA-1.5 13B in understanding performance, achieving a 7.2% average improvement. For image reconstruction, we achieve a strong FID score of 0.63 at 384x384 resolution. Moreover, TokenFlow establishes state-of-the-art performance in autoregressive image generation with a GenEval score of 0.55 at 256x256 resolution, achieving comparable results to SDXL. Our code and models are released at https://github.com/ByteFlow-AI/TokenFlow.

The adoption of Vision Transformers (ViTs) in resource-constrained applications necessitates improvements in inference throughput. To this end several token pruning and merging approaches have been proposed that improve efficiency by successively reducing the number of tokens. However, it remains an open problem to design a token reduction method that is fast, maintains high performance, and is applicable to various vision tasks. In this work, we present a token pruner that uses auxiliary prediction heads that learn to select tokens end-to-end based on task relevance. These auxiliary heads can be removed after training, leading to throughput close to that of a random pruner. We evaluate our method on image classification, semantic segmentation, object detection, and instance segmentation, and show speedups of 1.5 to 4x with small drops in performance. As a best case, on the ADE20k semantic segmentation benchmark, we observe a 2x speedup relative to the no-pruning baseline, with a negligible performance penalty of 0.1 median mIoU across 5 seeds.

Diffusion Transformers have demonstrated remarkable capabilities in image generation but often come with excessive parameterization, resulting in considerable inference overhead in real-world applications. In this work, we present TinyFusion, a depth pruning method designed to remove redundant layers from diffusion transformers via end-to-end learning. The core principle of our approach is to create a pruned model with high recoverability, allowing it to regain strong performance after fine-tuning. To accomplish this, we introduce a differentiable sampling technique to make pruning learnable, paired with a co-optimized parameter to simulate future fine-tuning. While prior works focus on minimizing loss or error after pruning, our method explicitly models and optimizes the post-fine-tuning performance of pruned models. Experimental results indicate that this learnable paradigm offers substantial benefits for layer pruning of diffusion transformers, surpassing existing importance-based and error-based methods. Additionally, TinyFusion exhibits strong generalization across diverse architectures, such as DiTs, MARs, and SiTs. Experiments with DiT-XL show that TinyFusion can craft a shallow diffusion transformer at less than 7% of the pre-training cost, achieving a 2xspeedup with an FID score of 2.86, outperforming competitors with comparable efficiency. Code is available at https://github.com/VainF/TinyFusion

As a fundamental backbone for video generation, diffusion models are challenged by low inference speed due to the sequential nature of denoising.Previous methods speed up the models by caching and reusing model outputs at uniformly selected timesteps.However, such a strategy neglects the fact that differences among model outputs are not uniform across timesteps, which hinders selecting the appropriate model outputs to cache, leading to a poor balance between inference efficiency and visual quality.In this study, we introduce Timestep Embedding Aware Cache (TeaCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps.Rather than directly using the time-consuming model outputs, TeaCache focuses on model inputs, which have a strong correlation with the modeloutputs while incurring negligible computational cost.TeaCache first modulates the noisy inputs using the timestep embeddings to ensure their differences better approximating those of model outputs. TeaCache then introduces a rescaling strategy to refine the estimated differences and utilizes them to indicate output caching.Experiments show that TeaCache achieves up to 4.41xacceleration over Open-Sora-Plan with negligible (-0.07% Vbench score) degradation of visual quality.

Video frame interpolation (VFI) that leverages the bio-inspired event cameras as guidance has recently shown better performance and memory efficiency than the frame-based methods, thanks to the event cameras' advantages, such as high temporal resolution. A hurdle for event-based VFI is how to effectively deal with non-linear motion, caused by the dynamic changes in motion direction and speed within the scene. Existing methods either use events to estimate sparse optical flow or fuse events with image features to estimate dense optical flow. Unfortunately, motion errors often degrade the VFI quality as the continuous motion cues from events do not align with the dense spatial information of images in the temporal dimension. In this paper, we find that object motion is continuous in space, tracking local regions over continuous time enables more accurate identification of spatiotemporal feature correlations. In light of this, we propose a novel continuous point tracking-based VFI framework, named TimeTracker. Specifically, we first design a Scene-Aware Region Segmentation (SARS) module to divide the scene into similar patches. Then, a Continuous Trajectory guided Motion Estimation (CTME) module is proposed to track the continuous motion trajectory of each patch through events. Finally, intermediate frames at any given time are generated through global motion optimization and frame refinement. Moreover, we collect a real-world dataset that features fast non-linear motion. Extensive experiments show that our method outperforms prior arts in both motion estimation and frame interpolation quality.