ICCV 2025 论文列表

Diffusion probabilistic models have shown significant progress in video generation; however, their computational efficiency is limited by the large number of sampling steps required. Reducing sampling steps often compromises video quality or generation diversity. In this work, we introduce a distillation method that combines variational score distillation and consistency distillation to achieve few-step video generation, maintaining both high quality and diversity. We also propose a latent reward model fine-tuning approach to further enhance video generation performance according to any specified reward metric. This approach reduces memory usage and does not require the reward to be differentiable. Our method demonstrates state-of-the-art performance in few-step generation for 10-second videos (128 frames at 12 FPS). The distilled student model achieves a score of 82.57 on VBench, surpassing the teacher model as well as baseline models Gen-3, T2V-Turbo, and Kling. One-step distillation accelerates the teacher model's diffusion sampling by up to 278.6 times, enabling near real-time generation. Human evaluations further validate the superior performance of our 4-step student models compared to teacher model using 50-step DDIM sampling.

With recent advances in Multimodal Large Language Models (MLLMs), grounding and referring capabilities have gained increasing attention for achieving detailed understanding and flexible user interaction. However, these capabilities still remain underdeveloped in visual document understanding due to the scarcity of fine-grained datasets and comprehensive benchmarks. To fill this gap, we propose the **DO**cument **G**rounding and **R**eferring data engine (**DOGR-Engine**), which generates two types of high-quality fine-grained document data: (1) multi-granular parsing data to improve text localization and recognition, and (2) instruction-tuning data to activate MLLMs' grounding and referring capabilities in dialogue and reasoning. Using the DOGR-Engine, we construct **DOGR-Bench**, a benchmark covering seven grounding and referring tasks across three document types (chart, poster, and PDF document), offering a comprehensive evaluation of fine-grained document understanding. Leveraging the generated data, we further develop **DOGR**, a strong baseline model that excels in text localization and recognition, while precisely grounds and refers to key textual information during conversation and reasoning, thereby advancing document understanding to a finer granularity and enable flexible interaction paradigms. Our code, data, and model are open-sourced at https://github.com/zyinan99/DOGR.

Recent methods have shown that pre-trained diffusion models can be fine-tuned to enable generative inverse rendering by learning image-conditioned noise-to-intrinsic mapping. Despite their remarkable progress, they struggle to robustly produce high-quality results as the noise-to-intrinsic paradigm essentially utilizes noisy images with deteriorated structure and appearance for intrinsic prediction, while it is common knowledge that structure and appearance information in an image are crucial for inverse rendering. To address this issue, we present DNF-Intrinsic, a robust yet efficient inverse rendering approach fine-tuned from a pre-trained diffusion model, where we propose to take the source image rather than Gaussian noise as input to directly predict deterministic intrinsic properties via flow matching. Moreover, we design a generative renderer to constrain that the predicted intrinsic properties are physically faithful to the source image. Experiments on both synthetic and real-world datasets show that our method clearly outperforms existing state-of-the-art methods. Our code is available at https://github.com/OnlyZZZZ/DNF-Intrinsic.

Recent probabilistic methods for 3D triangular meshes capture diverse shapes by differentiable mesh connectivity, but face high computational costs with increased shape details. We introduce a new differentiable mesh processing method that addresses this challenge and efficiently handles meshes with intricate structures. Our method reduces time complexity from O(N) to O(log N) and requires significantly less memory than previous approaches. Building on this innovation, we present a reconstruction algorithm capable of generating complex 2D and 3D shapes from point clouds or multi-view images. Visit https://sonsang.github.io/dmesh2-project for source code and supplementary material.

Diffusion models have achieved remarkable success in image generation but come with significant computational costs, posing challenges for deployment in resource-constrained environments. Recent post-training quantization (PTQ) methods have attempted to mitigate this issue by focusing on the iterative nature of diffusion models. However, these approaches often overlook outliers, leading to degraded performance at low bit-widths. In this paper, we propose a DMQ which combines Learned Equivalent Scaling (LES) and channel-wise Power-of-Two Scaling (PTS) to effectively address these challenges. Learned Equivalent Scaling optimizes channel-wise scaling factors to redistribute quantization difficulty between weights and activations, reducing overall quantization error. Recognizing that early denoising steps, despite having small quantization errors, crucially impact the final output due to error accumulation, we incorporate an adaptive timestep weighting scheme to prioritize these critical steps during learning. Furthermore, identifying that layers such as skip connections exhibit high inter-channel variance, we introduce channel-wise Power-of-Two Scaling for activations. To ensure robust selection of PTS factors even with small calibration set, we introduce a voting algorithm that enhances reliability. Extensive experiments demonstrate that our method significantly outperforms existing works, especially at low bit-widths such as W4A6 (4-bit weight, 6-bit activation) and W4A8, maintaining high image generation quality and model stability. The code is available at https://github.com/LeeDongYeun/dmq.

In this paper, we address the problem of small object detection (SOD) by introducing our novel approach - Dynamically Multiplexed Expanded Features Set (DM-EFS) form. Detecting small objects is challenging as they usually suffer from inadequate feature representation. Hence, to address this, we propose the Expanded Features Set (EFS) form - a simple yet effective idea to improve the feature representation of small objects by utilizing the untapped higher resolution features from the shallower layers of the backbone module. We observe that the EFS form improves the SOD performance. However, due to processing of additional features, it has a higher computational cost which reduces inference efficiency. Hence, to address this, we propose Dynamic Feature Multiplexing (DFM) - a novel design that optimizes the usage of the EFS form during inference by dynamically multiplexing it to create our aforementioned DM-EFS form. Since our DM-EFS form is a multiplexed (or subsampled) optimal version of the EFS form, it improves the SOD performance like the EFS form but with a lower computational cost. Extensive experiments confirm the efficacy of our DM-EFS approach. Integrated with YOLOv7 base model, our DM-EFS achieves state-of-the art results on diverse SOD datasets outperforming the base model and SOD baselines, with on-par or even better inference efficiency.

Diffusion Transformer (DiT)-based generation models have achieved remarkable success in video generation. However, their inherent computational demands pose significant efficiency challenges. In this paper, we exploit the inherent temporal non-uniformity of real-world videos, and observe that videos exhibit dynamic information density, with high-motion segments demanding greater detail preservation than static scenes. Inspired by this temporal non-uniformity, we propose DLFR-Gen, a training-free approach for Diffusion-based Video Generation with Dynamic Latent Frame Rate. DLFR-Gen adaptively adjusts the number of elements in latent space based on the motion frequency of the latent space content, using fewer tokens for low-frequency segments while preserving detail in high-frequency segments. Specifically, our key contributions are: (1) A dynamic frame rate scheduler for DiT video generation that adaptively assigns frame rates for video segments. (2) A novel latent-space frame merging method to align latent representations with their denoised counterparts before merging those redundant in low-resolution space. (3) A preference analysis of Rotary Positional Embeddings (RoPE) across DiT layers, informing a tailored RoPE strategy optimized for semantic and local information capture. Experiments show that DLFR-Gen can achieve a speedup up to 3x for video generation with minimal quality degradation.

Recent studies in extreme image compression have achieved remarkable performance by compressing the tokens from generative tokenizers. However, these methods often prioritize clustering common semantics within the dataset, while overlooking the diverse details of individual objects. Consequently, this results in suboptimal reconstruction fidelity, especially at low bitrates. To address this issue, we introduce a Dual-generative Latent Fusion (DLF) paradigm. DLF decomposes the latent into semantic and detail elements, compressing them through two distinct branches. The semantic branch clusters high-level information into compact tokens, while the detail branch encodes perceptually critical details to enhance the overall fidelity. Additionally, we propose a cross-branch interactive design to reduce redundancy between the two branches, thereby minimizing the overall bit cost. Experimental results demonstrate the impressive reconstruction quality of DLF even below 0.01 bits per pixel (bpp). On the CLIC2020 test set, our method achieves bitrate savings of up to 27.93% on LPIPS and 53.55% on DISTS compared to MS-ILLM. Furthermore, DLF surpasses recent diffusion-based codecs in visual fidelity while maintaining a comparable level of generative realism. Project: https://dlfcodec.github.io/

Building on the success of diffusion models in image generation and editing, video editing has recently gained substantial attention. However, maintaining temporal consistency and motion alignment still remains challenging. To address these issues, this paper proposes DINO-guided Video Editing (DIVE), a framework designed to facilitate subject-driven editing in source videos conditioned on either target text prompts or reference images with specific identities. The core of DIVE lies in leveraging the powerful semantic features extracted from a pretrained DINOv2 model as implicit correspondences to guide the editing process. Specifically, to ensure temporal motion consistency, DIVE employs DINO features to align with the motion trajectory of the source video. For precise subject editing, DIVE incorporates the DINO features of reference images into a pretrained text-to-image model to learn Low-Rank Adaptations (LoRAs), effectively registering the target subject's identity. Extensive experiments on diverse real-world videos demonstrate that our framework can achieve high-quality editing results with robust motion consistency, highlighting the potential of DINO to contribute to video editing.

Deep neural networks have demonstrated remarkable success across numerous tasks, yet they remain vulnerable to Trojan (backdoor) attacks, raising serious concerns about their safety in real-world mission-critical applications. A common countermeasure is trigger inversion -- reconstructing malicious "shortcut" patterns (triggers) inserted by an adversary during training. Current trigger-inversion methods typically search the full pixel space under specific assumptions but offer no assurances that the estimated trigger is more than an adversarial perturbation that flips the model output. Here, we propose a data-free, zero-shot trigger-inversion strategy that restricts the search space while avoiding strong assumptions on trigger appearance. Specifically, we incorporate a diffusion-based generator guided by the target classifier; through iterative generation, we produce candidate triggers that align with the internal representations the model relies on for malicious behavior. Empirical evaluations, both quantitative and qualitative, show that our approach reconstructs triggers that effectively distinguish clean versus Trojaned models. DISTIL surpasses alternative methods by high margins, achieving up to 7.1% higher accuracy on the BackdoorBench dataset and a 9.4% improvement on trojaned object detection model scanning, offering a promising new direction for reliable backdoor defense without reliance on extensive data or strong prior assumptions about triggers. The code is available at https://github.com/AdaptiveMotorControlLab/DISTIL.

Resolving closely-spaced small targets in dense clusters presents a significant challenge in infrared imaging, as the overlapping signals hinder precise determination of their quantity, sub-pixel positions, and radiation intensities. While deep learning has advanced the field of infrared small target detection, its application to closely-spaced infrared small targets has not yet been explored. This gap exists primarily due to the complexity of separating superimposed characteristics and the lack of an open-source infrastructure. In this work, we propose the Dynamic Iterative Shrinkage Thresholding Network (DISTA-Net), which reconceptualizes traditional sparse reconstruction within a dynamic framework. DISTA-Net adaptively generates convolution weights and thresholding parameters to tailor the reconstruction process in real time. To the best of our knowledge, DISTA-Net is the first deep learning model designed specifically for the unmixing of closely-spaced infrared small targets, achieving superior sub-pixel detection accuracy. Moreover, we have established the first open-source ecosystem to foster further research in this field. This ecosystem comprises three key components: (1) CSIST-100K, a publicly available benchmark dataset; (2) CSO-mAP, a custom evaluation metric for sub-pixel detection; and (3) GrokCSO, an open-source toolkit featuring DISTA-Net and other state-of-the-art models, available at https://github.com/GrokCV/GrokCSO.

We introduce DIP, a novel unsupervised post-training method designed to enhance dense representations in large-scale pretrained vision encoders for in-context scene understanding. Unlike prior approaches using complex self-distillation architectures, our method trains the vision encoder using pseudo-tasks that simulate downstream in-context scenarios, inspired by meta-learning principles. To enable post-training on unlabeled data, we propose an automatic mechanism for generating in-context tasks that combines a pretrained diffusion model and the vision encoder. DI"P is simple, unsupervised, and computationally efficient, requiring under 9 hours on a single A100 GPU. By learning dense representations through pseudo in-context tasks, it achieves strong performance across a variety of downstream real-world in-context scene understanding tasks. It outperforms both the initial vision encoder and prior methods, offering a practical and effective solution for improving dense representations.

We present DIMO, a generative approach capable of generating diverse 3D motions for arbitrary objects from a single image. The core idea of our work is to leverage the rich priors in well-trained video models to extract the common motion patterns and then embed them into a shared low-dimensional latent space. Specifically, we first generate multiple videos of the same object with diverse motions. We then embed each motion into a latent vector and train a shared motion decoder to learn the distribution of motions represented by a structured and compact motion representation, i.e., neural key point trajectories. The canonical 3D Gaussians are then driven by these key points and fused to model the geometry and appearance. During inference time with learned latent space, we can instantly sample diverse 3D motions in a single-forward pass and support several interesting applications including 3D motion interpolation and language-guided motion generation.

Recent advances in text-to-image (T2I) models have achieved impressive quality and consistency. However, this has come at the cost of representation diversity. While automatic evaluation methods exist for benchmarking model diversity, they either require reference image datasets or lack specificity about the kind of diversity measured, limiting their adaptability and interpretability. To address this gap, we introduce the Does-it/Can-it framework, DIM-CIM, a reference-free measurement of default-mode diversity ("Does" the model generate images with expected attributes?) and generalization capacity ("Can" the model generate diverse attributes for a particular concept?). We construct the COCO-DIMCIM benchmark, which is seeded with COCO concepts and captions and augmented by a large language model. With COCO-DIMCIM, we find that widely-used models improve in generalization at the cost of default-mode diversity when scaling from 1.5B to 8.1B parameters. DIMCIM also identifies fine-grained failure cases, such as attributes that are generated with generic prompts but are rarely generated when explicitly requested. Finally, we use DIMCIM to evaluate the training data of a T2I model and observe a correlation of 0.85 between diversity in training images and default-mode diversity. Our work provides a flexible and interpretable framework for assessing T2I model diversity and generalization, enabling a more comprehensive understanding of model performance.

Recent Training-Free Open-Vocabulary Semantic Segmentation (TF-OVSS) leverages a pre-training vision-language model to segment images from open-set visual concepts without training and fine-tuning. The key of TF-OVSS is to improve the local spatial representation of CLIP by leveraging self-correlation maps, thus preserving its zero-sample capability and achieving open understanding. However, most TF-OVSS methods utilize the Multi-Head Self-Attention (MHSA) mechanism to generate self-correlation maps, neglecting the diversity among multiple heads. In this paper, we explore the diversity of MHSA, revealing that the contributions of single-head attention to the final results are varied and redundant. To address this issue, we introduce DIH-CLIP, a training-free CLIP model for open-vocabulary semantic segmentation. Specifically, we propose a Selective Head Attention (SHA) to replace the traditional MHSA in CLIP, which contains two key designs: (1) evaluating the diversity of multi-head attention via calculating information entropy scores of per head attention map and removing the redundant attention head with threshold; (2) transferring the local representation of single-head attention to the global CLIP feature to enhance the local spatial representation capability of CLIP. Furthermore, we embed SHA into the middle layers of CLIP to extract the plentiful details. Experiments on six benchmark datasets demonstrate the effectiveness of DIH-CLIP.

Mixture-of-Experts-based (MoE-based) diffusion models demonstrate remarkable scalability in high-fidelity image generation, yet their reliance on expert parallelism introduces critical communication bottlenecks. State-of-the-art methods alleviate such overhead in parallel diffusion inference through computation-communication overlapping, termed displaced parallelism. However, we identify that these techniques induce severe staleness --- the usage of outdated activations from previous timesteps that significantly degrades quality, especially in expert-parallel scenarios. We tackle this fundamental tension and propose DICE, a staleness-centric optimization framework with a three-fold approach: (1) Interweaved Parallelism introduces staggered pipelines, effectively halving step-level staleness for free; (2) Selective Synchronization operates at layer-level and protects layers vulnerable to staled activations; and (3) Conditional Communication, a token-level, training-free method that dynamically adjusts communication frequency based on token importance. Together, these strategies effectively reduce staleness, achieving 1.26x speedup with minimal quality degradation. Empirical results establish DICE as an effective and scalable solution. Our code is available at https://github.com/Cobalt-27/DICE

Diffusion models have shown to be strong representation learners, showcasing state-of-the-art performance across multiple domains. Aside from accelerated sampling, DDIM also enables the inversion of real images back to their latent codes. A direct inheriting application of this inversion operation is real image editing, where the inversion yields latent trajectories to be utilized during the synthesis of the edited image. Unfortunately, this practical tool has enabled malicious users to freely synthesize misinformative or deepfake contents with greater ease, which promotes the spread of unethical and abusive, as well as privacy-, and copyright-infringing contents. While defensive algorithms such as AdvDM and Photoguard have been shown to disrupt the diffusion process on these images, the misalignment between their objectives and the iterative denoising trajectory at test time results in weak disruptive performance. In this work, we present the DDIM Inversion Attack (DIA) that attacks the integrated DDIM trajectory path. Our results support the effective disruption, surpassing previous defensive methods across various editing methods. We believe that our frameworks and results can provide practical defense methods against the malicious use of AI for both the industry and the research community. Our code is available here: https://anonymous.4open.science/r/DIA-13419/.

Human-centric generative models are becoming increasingly popular, giving rise to various innovative tools and applications, such as talking face videos conditioned on text or audio prompts. The core of these capabilities lies in powerful pre-trained foundation models, trained on large-scale, high-quality datasets. However, many advanced methods rely on in-house data subject to various constraints, and other current studies fail to generate high-resolution face videos, which is mainly attributed to the significant lack of large-scale, high-quality face video datasets. In this paper, we introduce a human face video dataset, DH-FaceVid-1K. Our collection spans 1,200 hours in total, encompassing 270,043 video clips from over 20,000 individuals. Each sample includes corresponding speech audio, facial keypoints, and text annotations. Compared to other publicly available datasets, ours distinguishes itself through its multi-ethnic coverage and high-quality, comprehensive individual attributes. We establish multiple face video generation models supporting tasks such as text-to-video and image-to-video generation. In addition, we develop comprehensive benchmarks to validate the scaling law when using different proportions of the proposed dataset. Our primary aim is to contribute a face video dataset, particularly addressing the underrepresentation of Asian faces in existing curated datasets and thereby enriching the global spectrum of face-centric data and mitigating demographic biases. Project Page: https://luna-ai-lab.github.io/DH-FaceVid-1K/.

In this work, we investigate the generation of high-fidelity, audio-driven 3D Gaussian talking heads from monocular videos. We present DGTalker, an innovative framework designed for real-time, high-fidelity, and 3D-aware talking head synthesis. By leveraging Gaussian generative priors and treating the task as a latent space navigation problem, our method effectively alleviates the lack of 3D information and the low-quality detail reconstruction caused by the absence of structure priors in monocular videos, which is a longstanding challenge in existing 3DGS-based approaches. To ensure precise lip synchronization and nuanced expression control, we propose a disentangled latent space navigation method that independently models lip motion and talking expressions. Additionally, we introduce an effective masked cross-view supervision strategy to enable robust learning within the disentangled framework. We conduct extensive experiments and demonstrate that DGTalker surpasses current state-of-the-art methods in visual quality, motion accuracy, and controllability.