CVPR 2025 Papers

Movie dubbing describes the process of transforming a script into speech that aligns temporally and emotionally with a given movie clip while exemplifying the speaker's voice demonstrated in a short reference audio clip. This task demands the model bridge character performances and complicated prosody structures to build a high-quality video-synchronized dubbing track. The limited scale of movie dubbing datasets, along with the background noise inherent in audio data, hinder the acoustic modeling performance of trained models. To address these issues, we propose an acoustic-prosody disentangled two-stage method to achieve high-quality dubbing generation with precise prosody alignment. First, we propose a prosody-enhanced acoustic pre-training to develop robust acoustic modeling capabilities. Then, we freeze the pre-trained acoustic system and design a disentangled framework to model prosodic text features and dubbing style while maintaining acoustic quality. Additionally, we incorporate an in-domain emotion analysis module to reduce the impact of visual domain shifts across different movies, thereby enhancing emotion-prosody alignment. Extensive experiments show that our method performs favorably against the state-of-the-art models on two primary benchmarks. The project is available at https://zzdoog.github.io/ProDubber/.

Precise Event Spotting (PES) aims to identify events and their class from long, untrimmed videos, particularly in sports. The main objective of PES is to detect the event at the exact moment it occurs. Existing methods mainly rely on features from a large pre-trained network, which may not be ideal for the task. Furthermore, these methods overlook the issue of imbalanced event class distribution present in the data, negatively impacting performance in challenging scenarios. This paper demonstrates that an appropriately designed network, trained end-to-end, can outperform state-of-the-art (SOTA) methods. Particularly, we propose a network with a convolutional spatial-temporal feature extractor enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and a long-range temporal module. The ASTRM enhances the features with spatio-temporal information. Meanwhile, the long-range temporal module helps extract global context from the data by modeling long-range dependencies. To address the class imbalance issue, we introduce the Soft Instance Contrastive (Soft-IC) loss that promotes feature compactness and class separation. Extensive experiments show that the proposed method is efficient and outperforms the SOTA methods, specifically in more challenging settings.

We study the challenging problem of estimating the relative pose of three calibrated cameras from four point correspondences. We propose novel efficient solutions to this problem that are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views. We model this geometry either as an affine or a fully perspective geometry estimated using one additional approximate correspondence. We generate such an approximate correspondence using a very simple and efficient strategy, where the new point is the mean point of three corresponding input points. The new solvers are efficient and easy to implement, since they are based on existing efficient minimal solvers, i.e., the 4-point affine fundamental matrix, the well-known 5-point relative pose solver, and the \texttt P3P solver. Extensive experiments on real data show that the proposed solvers, when properly coupled with local optimization, achieve state-of-the-art results, with the novel solver based on approximate mean-point correspondences being more robust and accurate than the affine-based solver.

We propose a training-free approach to 3D editing that enables the editing of a single shape and the reconstruction of a mesh within a few minutes. Leveraging 4-view images, user-guided text prompts, and rough 2D masks, our method produces an edited 3D mesh that aligns with the prompt. For this, our approach performs synchronized multi-view image editing in 2D. However, targeted regions to be edited are ambiguous due to projection from 3D to 2D. To ensure precise editing only in intended regions, we develop a 3D segmentation pipeline that detects edited areas in 3D space. Additionally, we introduce a merging algorithm to seamlessly integrate edited 3D regions with original input. Extensive experiments demonstrate the superiority of our method over previous approaches, enabling fast, high-quality editing while preserving unintended regions.

As a widely adopted technique in data transmission, video compression effectively reduces the size of files, making it possible for real-time cloud computing. However, it comes at the cost of visual quality, posing challenges to the robustness of downstream vision models. In this work, we present a versatile codec-aware enhancement framework that reuses codec information to adaptively enhance videos under different compression settings, assisting various downstream vision tasks without introducing computation bottleneck. Specifically, the proposed codec-aware framework consists of a compression-aware adaptation (CAA) network that employs a hierarchical adaptation mechanism to estimate parameters of the frame-wise enhancement network, namely the bitstream-aware enhancement (BAE) network. The BAE network further leverages temporal and spatial priors embedded in the bitstream to effectively improve the quality of compressed input frames. Extensive experimental results demonstrate the superior quality enhancement performance of our framework over existing enhancement methods, as well as its versatility in assisting multiple downstream tasks on compressed videos as a plug-and-play module. Code and models are available at https://huimin-zeng.github.io/PnP-VCVE/.

Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD.

This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multi-view indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching/tracking for planar surface reconstruction. Furthermore, with the essential merits of plane-based representation coupled with CUDA-based implementation of planar splatting functions, PlanarSplatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available.

Diffusion prior-based methods have shown impressive results in real-world image super-resolution (SR). However, most existing methods entangle pixel-level and semantic-level SR objectives in the training process, struggling to balance pixel-wise fidelity and perceptual quality. Meanwhile, users have varying preferences on SR results, thus it is demanded to develop an adjustable SR model that can be tailored to different fidelity-perception preferences during inference without re-training. We present Pixel-level and Semantic-level Adjustable SR (PiSA-SR), which learns two LoRA modules upon the pre-trained stable-diffusion (SD) model to achieve improved and adjustable SR results. We first formulate the SD-based SR problem as learning the residual between the low-quality input and the high-quality output, then show that the learning objective can be decoupled into two distinct LoRA weight spaces: one is characterized by the l2-loss for pixel-level regression, and another is characterized by the LPIPS and classifier score distillation losses to extract semantic information from pre-trained classification and SD models. In its default setting, PiSA-SR can be performed in a single diffusion step, achieving leading real-world SR results in both quality and efficiency. By introducing two adjustable guidance scales on the two LoRA modules to control the strengths of pixel-wise fidelity and semantic-level details during inference, PiSA-SR can offer flexible SR results according to user preference without re-training. The source code of our method can be found at https://github.com/csslc/PiSA-SR.

RLHF techniques like DPO can significantly improve the generation quality of text-to-image diffusion models. However, these methods optimize for a single reward that aligns model generation with population-level preferences, neglecting the nuances of individual users' beliefs or values. This lack of personalization limits the efficacy of these models. To bridge this gap, we introduce PPD, a multi-reward optimization objective that aligns diffusion models with personalized preferences. With PPD, a diffusion model learns the individual preferences of a population of users in a few-shot way, enabling generalization to unseen users. Specifically, our approach (1) leverages a vision-language model (VLM) to extract personal preference embeddings from a small set of pairwise preference examples, and then (2) incorporates the embeddings into diffusion models through cross attention. Conditioning on user embeddings, the text-to-image models are fine-tuned with the DPO objective, simultaneously optimizing for alignment with the preferences of multiple users. Empirical results demonstrate that our method effectively optimizes for multiple reward functions and can interpolate between them during inference. In real-world user scenarios, with as few as four preference examples from a new user, our approach achieves an average win rate of 76% over Stable Cascade, generating images that more accurately reflect specific user preferences.

Finetuning-free personalized image generation can synthesize customized images without test-time finetuning, attracting wide research interest owing to its high efficiency. Current finetuning-free methods simply adopt a single training stage with a simple image reconstruction task, and they typically generate low-quality images inconsistent with the reference images during test-time. To mitigate this problem, inspired by the recent DPO (i.e., direct preference optimization) technique, this work proposes an additional training stage to improve the pre-trained personalized generation models. However, traditional DPO only determines the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because the generated images are commonly inconsistent with the reference images only in some local image patches. To tackle this problem, this work proposes PatchDPO that estimates the quality of image patches within each generated image and accordingly trains the model. To this end, PatchDPO first leverages the pre-trained vision models with a proposed self-supervised training method to estimate the patch quality. Next, PatchDPO adopts a weighted training approach to train the model with the estimated patch quality, which rewards the image patches with high quality while penalizing the image patches with low quality. Experiment results demonstrate that PatchDPO significantly improves the performance of multiple pre-trained personalized generation models, and achieves state-of-the-art performance on both single-object and multi-object personalized image generation. Our code is available at https://github.com/hqhQAQ/PatchDPO.

Recently, Depth Anything Models (DAMs) - a type of depth foundation models - have demonstrated impressive zero-shot capabilities across diverse perspective images. Despite its success, it remains an open question regarding DAMs' performance on panorama images that enjoy a large field-of-view (180x360) but suffer from spherical distortions. To address this gap, we conduct an empirical analysis to evaluate the performance of DAMs on panoramic images and identify their limitations. For this, we undertake comprehensive experiments to assess the performance of DAMs from three key factors: panoramic representations, 360 camera positions for capturing scenarios, and spherical spatial transformations. This way, we reveal some key findings, e.g., DAMs are sensitive to spatial transformations. We then propose a semi-supervised learning (SSL) framework to learn a panoramic DAM, dubbed PanDA. Under the umbrella of SSL, PanDA first learns a teacher model by fine-tuning DAM through joint training on synthetic indoor and outdoor panoramic datasets. Then, a student model is trained using large-scale unlabeled data, leveraging pseudo-labels generated by the teacher model. To enhance PanDA's generalization capability, Mobius transformation-based spatial augmentation (MTSA) is proposed to impose consistency regularization between the predicted depth maps from the original and spatially transformed ones. This subtly improves the student model's robustness to various spatial transformations, even under severe distortions. Extensive experiments demonstrate that PanDA exhibits remarkable zero-shot capability across diverse scenes, and outperforms the data-specific panoramic depth estimation methods on two popular real-world benchmarks.

Existing LVLM-based reasoning segmentation methods often suffer from imprecise segmentation results and hallucinations in their text responses. This paper introduces POPEN, a novel framework designed to address these issues and achieve improved results. POPEN includes a preference-based optimization method to finetune the LVLM, aligning it more closely with human preferences and thereby generating better text responses and segmentation results. Additionally, POPEN introduces a preference-based ensemble method for inference, which integrates multiple outputs from the LVLM using a preference-score-based attention mechanism for refinement. To better adapt to the segmentation task, we incorporate several task-specific designs in our POPEN framework, including a new approach for collecting segmentation preference data with a curriculum learning mechanism, and a novel preference optimization loss to refine the segmentation capability of the LVLM. Experiments demonstrate that our method achieves state-of-the-art performance in reasoning segmentation, exhibiting minimal hallucination in text responses and the highest segmentation accuracy compared to previous advanced methods like LISA and PixelLM.

The democratization of machine learning systems has made the process of fine-tuning accessible to practitioners, leading to a wide range of open-source models fine-tuned on specialized tasks and datasets. Recent work has proposed to merge such models to combine their functionalities. However, prior approaches are usually restricted to models that are fine-tuned from the same base model. Furthermore, the final merged model is typically required to be of the same size as the original models. In this work, we propose a new two-step algorithm to merge models---termed PLeaS---which relaxes these constraints.First, leveraging the Permutation symmetries inherent in the two models, PLeaS partially matches nodes in each layer by maximizing alignment. Next, PLeaS computes the weights of the merged model as a layer-wise Least Squares solution to minimize the approximation error between the features of the merged model and the permuted features of the original models. PLeaS allows a practitioner to merge two models sharing the same architecture into a single performant model of a desired size, even when the two original models are fine-tuned from different base models. We also demonstrate how our method can be extended to address a challenging scenario where no data is available from the fine-tuning domains. We demonstrate our method to merge ResNet and ViT models trained with shared and different label spaces, and show improvement over the state-of-the-art merging methods of up to 15 percentage points for the same target compute while merging models trained on DomainNet and fine-grained classification tasks.

Polarization cameras can capture multiple polarized images with different polarizer angles in a single shot, bringing convenience to polarization-based downstream tasks. However, their direct outputs are color-polarization filter array (CPFA) raw images, requiring demosaicing to reconstruct full-resolution, full-color polarized images; unfortunately, this necessary step introduces artifacts that make polarization-related parameters such as the degree of polarization (DoP) and angle of polarization (AoP) prone to error. Besides, limited by the hardware design, the resolution of a polarization camera is often much lower than that of a conventional RGB camera. Existing polarized image demosaicing (PID) methods are limited in that they cannot enhance resolution, while polarized image super-resolution (PISR) methods, though designed to obtain high-resolution (HR) polarized images from the demosaicing results, tend to retain or even amplify errors in the DoP and AoP introduced by demosaicing artifacts. In this paper, we propose PIDSR, a joint framework that performs complementary Polarized Image Demosaicing and Super-Resolution, showing the ability to robustly obtain high-quality HR polarized images with more accurate DoP and AoP from a CPFA raw image in a direct manner. Experiments show our PIDSR not only achieves state-of-the-art performance on both synthetic and real data, but also facilitates downstream tasks.

Visual Language Models require substantial computational resources for inference due to the additional input tokens needed to represent visual information. However, these visual tokens often contain redundant and unimportant information, resulting in an unnecessarily high number of tokens. To address this, we introduce PACT, a method that reduces inference time and memory usage by pruning irrelevant tokens and merging visually redundant ones at an early layer of the language model. Our approach uses a novel importance metric to identify unimportant tokens without relying on attention scores, making it compatible with FlashAttention. We also propose a novel clustering algorithm, called Distance Bounded Density Peak Clustering, which efficiently clusters visual tokens while constraining the distances between elements within a cluster by a predefined threshold. We demonstrate the effectiveness of PACT through extensive experiments.

Recently, virtual staining has emerged as a promising alternative to revolutionize histological staining by digitally generating stains. However, most existing methods suffer from the curse of staining unreality and unreliability. In this paper, we propose the Orthogonal Decoupling Alignment Generative Adversarial Network (ODA-GAN) for unpaired virtual immunohistochemistry (IHC) staining. Our approach is based on the assumption that an image consists of IHC staining-related features, which influence staining distribution and intensity, and staining-unrelated features, such as tissue morphology. Leveraging a pathology foundation model, we first develop a weakly-supervised segmentation pipeline as an alternative to expert annotations. We introduce an Orthogonal MLP (O-MLP) module to project image features into an orthogonal space, decoupling them into staining-related and unrelated components. Additionally, we propose a Dual-stream PatchNCE (DPNCE) loss to resolve contrastive learning contradictions in the staining-related space, thereby enhancing staining accuracy. To further improve realism, we introduce a Multi-layer Domain Alignment (MDA) module to bridge the domain gap between generated and real IHC images. Evaluations on three benchmark datasets show that our ODA-GAN reaches state-of-the-art (SOTA) performance. Our source code is available at https://github.com/ittong/ODA-GAN.

Existing domain generalization methods for LiDAR semantic segmentation under adverse weather struggle to accurately predict "things" categories compared to "stuff" categories. In typical driving scenes, "things" categories can be dynamic and associated with higher collision risks, making them crucial for safe navigation and planning. Recognizing the importance of "things" categories, we identify their performance drop as a serious bottleneck in existing approaches. We observed that adverse weather induces degradation of semantic-level features and both corruption of local features, leading to a misprediction of "things" as "stuff". To mitigate these corruptions, we suggests our method, NTN - segmeNt Things for No-accident. To address semantic-level feature corruption, we bind each point feature to its superclass, preventing the misprediction of things classes into visually dissimilar categories. Additionally, to enhance robustness against local corruption caused by adverse weather, we define each LiDAR beam as a local region and propose a regularization term that aligns the clean data with its corrupted counterpart in feature space. NTN achieves state-of-the-art performance with a +2.6 mIoU gain on the SemanticKITTI-to-SemanticSTF benchmark and +7.9 mIoU on the SemanticPOSS-to-SemanticSTF benchmark. Notably, NTN achieves a +4.8 and +7.9 mIoU improvement on "things" classes, respectively, highlighting its effectiveness.

Optical coherence tomography angiography (OCTA) shows its great importance in imaging microvascular networks by providing accurate 3D imaging of blood vessels, but it relies upon specialized sensors and expensive devices. For this reason, previous works show the potential to translate the readily available 3D Optical Coherence Tomography (OCT) images into 3D OCTA images. However, existing OCTA translation methods directly learn the mapping from the OCT domain to the OCTA domain in continuous and infinite space with guidance from only a single view, i.e., the OCTA project map, resulting in suboptimal results. To this end, we propose the multi-view Tri-alignment framework for OCT to OCTA 3D image translation in discrete and finite space, named MuTri. In the first stage, we pre-train two vector-quantized variational auto-encoder (VQVAE) by reconstructing 3D OCT and 3D OCTA data, providing semantic prior for subsequent multi-view guidances. In the second stage, our multi-view tri-alignment facilitates another VQVAE model to learn the mapping from the OCT domain to the OCTA domain in discrete and finite space. Specifically, a contrastive-inspired semantic alignment is proposed to maximize the mutual information with the pre-trained models from OCT and OCTA views, to facilitate codebook learning. Meanwhile, a vessel structure alignment is proposed to minimize the structure discrepancy with the pre-trained models from the OCTA project map view, benefiting from learning the detailed vessel structure information. We also collect the first large-scale dataset, namely, OCTA2024, which contains a pair of OCT and OCTA volumes from 846 subjects. Our codes and datasets are available at: https://github.com/xmed-lab/MuTri.