ICCV 2025 Papers

The integration of Rotary Position Embedding (RoPE) in Multimodal Diffusion Transformer (MMDiT) has significantly enhanced text-to-image generation quality. However, the fundamental reliance of self-attention layers on positional embedding versus query-key similarity during generation remains an intriguing question. We present the first mechanistic analysis of RoPE-based MMDiT models (e.g., FLUX), introducing an automated probing strategy that disentangles positional information versus content dependencies by strategically manipulating RoPE during generation. Our analysis reveals distinct dependency patterns that do not straightforwardly correlate with depth, offering new insights into the layer-specific roles in RoPE-based MMDiT. Based on these findings, we propose a training-free, task-specific image editing framework that categorizes editing tasks into three types: position-dependent editing (e.g., object addition), content similarity-dependent editing (e.g., non-rigid editing), and region-preserved editing (e.g., background replacement). For each type, we design tailored key-value injection strategies based on the characteristics of the editing task. Extensive qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches, particularly in preserving original semantic content and achieving seamless modifications.

Generating harmonic and diverse human motions from music signals, especially for a group of dancers, is a practical yet challenging task in virtual avatar creation. Existing methods merely model a fixed number of dancers, lacking the flexibility for arbitrary individuals. To fulfill this goal, we propose a novel unified framework, FreeDance. Considering the plausibility of generating arbitrary dancers while preserving the diverse dynamics of multiple individuals, we build the framework upon collaborative masked token modeling in a 2D discrete space. In particular, we devise a Cross-modality Residual Alignment Module (CRAM) to diversify the movement of each individual and intensify its alignment with music. CRAM captures the spatial motion deformation of each dancer using residual learning and integrates it with rhythmic representation to reinforce the intrinsic connection. Moreover, recognizing the requirement of interactive coordination, we design a Temporal Interaction Module (TIM). Benefiting from masked 2D motion tokens, TIM effectively models the temporal correlation between current individuals w.r.t. neighboring dancers as interaction guidance to foster stronger inter-dancer dependencies. Extensive experiments demonstrate that our approach generates harmonic group dance with any number of individuals, outperforming state-of-the-art methods adapting number-fixed counterparts. Code is available at https://github.com/Tsukasane/FreeDance.

Domain Adaptation (DA) for dense prediction tasks is an important topic, which enhances the dense prediction model's performance when tested on its unseen domain. Recently, with the development of Diffusion-based Dense Prediction (DDP) models, the exploration of DA designs tailored to this framework is worth exploring, since the diffusion model is effective in modeling the distribution transformation that comprises domain information. In this work, we propose a training-free mechanism for DDP frameworks, endowing them with DA capabilities. Our motivation arises from the observation that the exposure bias (e.g., noise statistics bias) in diffusion brings domain shift, and different domains in conditions of DDP models can also be effectively captured by the noise prediction statistics. Based on this, we propose a training-free Domain Noise Alignment (DNA) approach, which alleviates the variations of noise statistics to domain changes during the diffusion sampling process, thereby achieving domain adaptation. Specifically, when the source domain is available, we directly adopt the DNA method to achieve domain adaptation by aligning the noise statistics of the target domain with those of the source domain. For the more challenging source-free DA, inspired by the observation that regions closer to the source domain exhibit higher confidence meeting variations of sampling noise, we utilize the statistics from the high-confidence regions progressively to guide the noise statistic adjustment during the sampling process. Notably, our method demonstrates the effectiveness of enhancing the DA capability of DDP models across four common dense prediction tasks. Code is available at \href https://github.com/xuhang07/FreeDNA https://github.com/xuhang07/FreeDNA .

In light of recent breakthroughs in text-to-image (T2I) generation, particularly with diffusion transformers (DiT), subject-driven technologies are increasingly being employed for high-fidelity customized production that preserves subject identity from reference inputs, enabling thrilling design workflows and engaging entertainment. Existing alternatives typically require either per-subject optimization via trainable text embeddings or training specialized encoders for subject feature extraction on large-scale datasets. Such dependencies on training procedures fundamentally constrain their practical applications. More importantly, current methodologies fail to fully leverage the inherent zero-shot potential of modern diffusion transformers (e.g., the Flux series) for authentic subject-driven synthesis. To bridge this gap, we propose FreeCus, a genuinely training-free framework that activates DiT's capabilities through three key innovations: 1) We introduce a pivotal attention sharing mechanism that captures the subject's layout integrity while preserving crucial editing flexibility. 2) Through a straightforward analysis of DiT's dynamic shifting, we propose an upgraded variant that significantly improves fine-grained feature extraction. 3) We further integrate advanced Multimodal Large Language Models (MLLMs) to enrich cross-modal semantic representations. Extensive experiments reflect that our method successfully unlocks DiT's zero-shot ability for consistent subject synthesis across diverse contexts, achieving state-of-the-art or comparable results compared to approaches that require additional training. Notably, our framework demonstrates seamless compatibility with existing inpainting pipelines and control modules, facilitating more compelling experiences. Our code is available at: https://github.com/Monalissaa/FreeCus.

We present Free4D, a novel tuning-free framework for 4D scene generation from a single image. Existing methods either focus on object-level generation, making scene-level generation infeasible, or rely on large-scale multi-view video datasets for expensive training, with limited generalization ability due to the scarcity of 4D scene data. In contrast, our key insight is to distill pre-trained foundation models for consistent 4D scene representation, which offers promising advantages such as efficiency and generalizability. 1) To achieve this, we first animate the input image using image-to-video diffusion models followed by 4D geometric structure initialization. 2) To turn this coarse structure into spatial-temporal consistent multi-view videos, we design an adaptive guidance mechanism with a point-guided denoising strategy for spatial consistency and a novel latent replacement strategy for temporal coherence. 3) To lift these generated observations into consistent 4D representation, we propose a modulation-based refinement to mitigate inconsistencies while fully leveraging the generated information. The resulting 4D representation enables real-time, controllable rendering, marking a significant advancement in single-image-based 4D scene generation.

Diffusion models have achieved impressive results in generative tasks for text-to-video (T2V) synthesis. However, achieving accurate text alignment in T2V generation remains challenging due to the complex temporal dependencies across frames. Existing reinforcement learning (RL)-based approaches to enhance text alignment often require differentiable reward functions trained for video, hindering their scalability and applicability. In this paper, we propose Free^2Guide, a novel gradient-free and training-free framework for aligning generated videos with text prompts. Specifically, leveraging principles from path integral control, Free^2Guide approximates guidance for diffusion models using non-differentiable reward functions, thereby enabling the integration of powerful black-box Large Vision-Language Models (LVLMs) as reward models. To enable image-trained LVLMs to assess text-to-video alignment, we leverage stitching between video frames and use system prompts to capture sequential attributions. Our framework supports the flexible ensembling of multiple reward models to synergistically enhance alignment without significant computational overhead. Experimental results confirm that Free^2Guide using image-trained LVLMs significantly improves text-to-video alignment, thereby enhancing the overall video quality. Our results and code are available at https://free2guide.github.io/

Conventional wisdom suggests that single-photon lidar (SPL) should operate in low-light conditions to minimize dead-time effects.Many methods have been developed to mitigate these effects in synchronous SPL systems. However, solutions for free-running SPL remain limited despite the advantage of reduced histogram distortion from dead times.To improve the accuracy of free-running SPL, we propose a computationally efficient joint maximum likelihood estimator of the signal flux, the background flux, and the depth, along with a complementary regularization framework that incorporates a learned point cloud score model as a prior.Simulations and experiments demonstrate that free-running SPL yields lower estimation errors than its synchronous counterpart under identical conditions, with our regularization further improving accuracy.

Video Multimodal Large Language Models (Video-MLLM) have achieved remarkable advancements in video understanding tasks. However, constrained by the context length limitation in the underlying LLMs, existing Video-MLLMs typically exhibit suboptimal performance on long video scenarios. To understand extended input frames, common solutions span token compression and streaming inference techniques, which sacrifice feature granularity or inference efficiency. Differently, to efficiently achieve comprehensive understanding of longer frame inputs, we draw ideas from MoE and propose a training-free approach Free-MoRef, which instantly multiplexes the context perception capabilities of Video-MLLMs within one inference pass. Specifically, Free-MoRef reconstructs the vision tokens into several short sequences as multi-references. Subsequently, we introduce MoRef-attention, which gathers clues from the multi-reference chunks in parallel to summarize unified query activations. After the shadow layers in LLMs, a reference fusion step is derived to compose a final mixed reasoning sequence with key tokens from parallel chunks, which compensates the cross-reference vision interactions that are neglected in MoRef-attention. By splitting and fusing the long vision token sequences, Free-MoRef achieves improved performance under much lower computing costs in reasoning multiplexed context length, demonstrating strong efficiency and effectiveness. Experiments on VideoMME, MLVU, LongVideoBench show that Free-MoRef achieves full perception of 2xto 8xlonger input frames without compression on a single A100 GPU while keeping instant responses, thereby bringing significant performance gains, even surpassing dedicatedly trained long-video-MLLMs. Codes are available at https://github.com/wkfdb/Free-MoRef

Controlling the movements of dynamic objects and the camera within generated videos is a meaningful yet challenging task. Due to the lack of datasets with comprehensive 6D pose annotations, existing text-to-video methods can not simultaneously control the motions of both camera and objects in 3D-aware manner, resulting in limited controllability over generated contents. To address this issue and facilitate the research in this field, we introduce a Synthetic Dataset for Free-Form Motion Control (SynFMC). The proposed SynFMC dataset includes diverse object and environment categories and covers various motion patterns according to specific rules, simulating common and complex real-world scenarios. The complete 6D pose information facilitates models learning to disentangle the motion effects from objects and the camera in a video. To provide precise 3D-aware motion control, we further propose a method trained on SynFMC, Free-Form Motion Control (FMC). FMC can control the 6D poses of objects and camera independently or simultaneously, producing high-fidelity videos. Moreover, it is compatible with various personalized text-to-image (T2I) models for different content styles. Extensive experiments demonstrate that the proposed FMC outperforms previous methods across multiple scenarios.

Interactive image editing allows users to modify images through visual interaction operations such as drawing, clicking, and dragging. Existing methods construct such supervision signals from videos, as they capture how objects change with various physical interactions. However, these models are usually built upon text-to-image diffusion models, so they either rely on slow optimization-based inference, or necessitate (i) massive training samples and (ii) an additional reference encoder to learn real-world dynamics and visual consistency. In this paper, we reformulate this task as an image-to-video generation problem, so that inherit powerful video diffusion priors to reduce training costs and ensure temporal consistency. Specifically, we introduce FramePainter as an efficient instantiation of this formulation. Initialized with Stable Video Diffusion, it only uses a lightweight sparse control encoder to inject editing signals. Considering the limitations of temporal attention in handling large motion between two frames, we further propose matching attention to enlarge the receptive field while encouraging dense correspondence between edited and source image tokens. We highlight the effectiveness and efficiency of FramePainter across various of editing signals: it domainantly outperforms previous state-of-the-art methods with far less training data, achieving highly seamless and coherent editing of images, e.g., automatically adjust the reflection of the cup. Moreover, FramePainter also exhibits exceptional generalization in scenarios not present in real-world videos, e.g., transform the clownfish into shark-like shape. Our code is available at https://github.com/YBYBZhang/FramePainter.

The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily prune tokens based on importance metrics, such as cumulative attention scores. However, even important tokens may exhibit high redundancy caused by similarity among adjacent video frames and repetitive visual elements. To address this limitation, we propose FrameFusion, a novel token reduction approach integrating similarity-based merging with importance-based pruning. We conduct a thorough study on token similarity characteristics, revealing three key insights: (1) spatially corresponding visual tokens between adjacent frames have higher cosine similarities compared to other token pairs; (2) high token similarities prominently decrease in deeper model layers; and (3) token similarity rankings are highly consistent across different layers. Guided by these observations, FrameFusion computes token similarities exclusively between corresponding visual tokens from adjacent frames, applies token merging at initial successive layers followed by pruning in deeper layers, and adopts a cascaded merging strategy to further enhance efficiency. We evaluate FrameFusion comprehensively across six diverse LVLMs, ranging from 2B to 72B parameters, using five video benchmarks encompassing video retrieval, question-answering, and spatial-temporal understanding tasks. Experiments show that FrameFusion reduces visual tokens by 70%, achieving 1.6-3.6x end-to-end speedups, with an average performance impact of less than 3%. Our code is available at: https://github.com/thu-nics/FrameFusion.

Despite the significant progress made by all-in-one models in universal image restoration, existing methods suffer from a generalization bottleneck in real-world scenarios, as they are mostly trained on small-scale synthetic datasets with limited degradations. Therefore, large-scale high-quality real-world training data is urgently needed to facilitate the emergence of foundation models for image restoration. To advance this field, we spare no effort in contributing a million-scale dataset with two notable advantages over existing training data: larger-scale real-world samples, and higher-diversity data types. By adjusting internal camera settings and external imaging conditions, we can capture aligned image pairs using our well-designed data acquisition system over multiple rounds and our data alignment criterion. Moreover, we propose a robust model, FoundIR, to better address a broader range of restoration tasks in real-world scenarios, taking a further step toward foundation models. Specifically, we first utilize a diffusion-based generalist model to remove degradations by learning the degradation-agnostic common representations from diverse inputs, where incremental learning strategy is adopted to better guide model training. To refine the model's restoration capability in complex scenarios, we introduce degradation-aware specialist models for achieving final high-quality results. Extensive experiments show the value of our dataset and the effectiveness of our method.

Federated Unlearning (FU) should satisfy three key requirements: a guarantee of data erasure, preservation of model utility, and reduction of unlearning time. Recent studies focus on identifying and modifying original model parameters relevant to unlearning data. While they can achieve faster unlearning, they degrade the model performance on remaining data or fail to forget unlearning data due to the difficulty in isolating specific parameters of the unlearning data. By revisiting the representation distribution of the optimal unlearning models (i.e., the retrained models), we observe that unlearning data tends to cluster within semantically related categories of remaining data. This inspired us to transform the distribution of unlearning data to fuse with similar categories in the remaining data for effective FU. Based on this insight, we propose a novel framework, named FUCRT, to achieve Federated Unlearning via Class-aware Representation Transformation. FUCRT consists of two key components: (1) a transformation class identification strategy (TCI) that leverages the original model to identify appropriate transformation classes for unlearning data, and (2) a targeted transformation learning process (TTL) with cross-class fusion mechanism to ensure effective and consistent transformation. Extensive experiments on four datasets demonstrate that FUCRT not only achieves 100% of data erasure but also outperforms state-of-the-art methods by an average of 2.96% and 3.78% in utility preservation under IID and Non-IID settings, respectively. Moreover, it reduces unlearning time by 19.13%~ 96.38%. Codes are at https://github.com/zhentian777/FUCRT.

The rise of generative models has raised concerns about image authenticity online, highlighting the urgent need for a detector that is (1) highly generalizable, capable of handling unseen forgery techniques, and (2) data-efficient, achieving optimal performance with minimal training data, enabling it to counter newly emerging forgery techniques effectively. To achieve this, we propose ForgeLens, a data-efficient, feature-guided framework that incorporates two lightweight designs to enable a frozen network to focus on forgery-specific features. First, we introduce the Weight-Shared Guidance Module (WSGM), which guides the extraction of forgery-specific features during training. Second, a forgery-aware feature integrator, FAFormer, is used to effectively integrate forgery information across multi-stage features. ForgeLens addresses a key limitation of previous frozen network-based methods, where general-purpose features extracted from large datasets often contain excessive forgery-irrelevant information. As a result, it achieves strong generalization and reaches optimal performance with minimal training data. Experimental results on 19 generative models, including both GANs and diffusion models, demonstrate improvements of 13.61% in Avg.Acc and 8.69% in Avg.AP over the base model. Notably, ForgeLens outperforms existing forgery detection methods, achieving state-of-the-art performance with just 1% of the training data.

The segmentation of forest LiDAR 3D point clouds, including both individual tree and semantic segmentation, is fundamental for advancing forest management and ecological research. However, current approaches often struggle with the complexity and variability of natural forest environments. We present ForestFormer3D, a new unified and end-to-end framework designed for precise individual tree and semantic segmentation. ForestFormer3D incorporates ISA-guided query point selection, a score-based block merging strategy during inference, and a one-to-many association mechanism for effective training. By combining these new components, our model achieves state-of-the-art performance for individual tree segmentation on the newly introduced FOR-instanceV2 dataset, which spans diverse forest types and regions. Additionally, ForestFormer3D generalizes well to unseen test sets (Wytham woods and LAUTx), showcasing its robustness across different forest conditions and sensor modalities. The FOR-instanceV2 dataset and the ForestFormer3D code are publicly available at https://bxiang233.github.io/FF3D/.

Motion forecasting for on-road traffic agents presents both a significant challenge and a critical necessity for ensuring safety in autonomous driving systems. In contrast to most existing data-driven approaches that directly predict future trajectories, we rethink this task from a planning perspective, advocating a "First Reasoning, Then Forecasting" strategy that explicitly incorporates behavior intentions as spatial guidance for trajectory prediction. To achieve this, we introduce an interpretable, reward-driven intention reasoner grounded in a novel query-centric Inverse Reinforcement Learning (IRL) scheme. Our method first encodes traffic agents and scene elements into a unified vectorized representation, then aggregates contextual features through a query-centric paradigm. This enables the derivation of a reward distribution, a compact yet informative representation of the target agent's behavior within the given scene context via IRL. Guided by this reward heuristic, we perform policy rollouts to reason about multiple plausible intentions, providing valuable priors for subsequent trajectory generation. Finally, we develop a hierarchical DETR-like decoder integrated with bidirectional selective state space models to produce accurate future trajectories along with their associated probabilities. Extensive experiments on the large-scale Argoverse and nuScenes motion forecasting datasets demonstrate that our approach significantly enhances trajectory prediction confidence, achieving highly competitive performance relative to state-of-the-art methods.

Recently, synthetic images have evolved incredibly realistic with the development of generative techniques. To avoid the spread of misinformation and identify synthetic content, research on synthetic image detection becomes urgent. Unfortunately, limited to the singular forensic perspective, existing methods struggle to explore sufficient traces encountered with diverse synthetic techniques. In response to this, we argue that different synthetic images encompass a variety of forensic traces, and utilizing multiple experts to explore traces from diverse perspectives will be beneficial. Accordingly, a novel detector with the Mixture of multiple forensic Experts is proposed, named Forensic-MoE. To integrate multiple experts and enhance the knowledge interaction, Forensic-MoE follows an adapter-backbone architecture. Specifically, multiple adapters trained on different synthetic images serve as the trace exploration experts, and they are uniformly integrated into a pretrained backbone model to learn the detection prior and encourage the expert interaction. By guiding multiple experts to align with each other and collaborate, Forensic-MoE can integrate comprehensive and discriminative detection traces from multiple perspectives. Moreover, for the discrimination improvement of each expert, a multi-stage structure is proposed for efficient trace perception, and a patch decentralization strategy is applied to encourage the model's attention on every local region. Code will be available in https://github.com/fangmq77/Forensic-MoE

Modeling the rotation of moving objects is a fundamental task in computer vision, yet SO(3) extrapolation still presents numerous challenges: (1) unknown quantities such as the moment of inertia complicate dynamics, (2) the presence of external forces and torques can lead to non-conservative kinematics, and (3) estimating evolving state trajectories under sparse, noisy observations requires robustness. We propose modeling trajectories of noisy pose estimates on the manifold of 3D rotations in a physically and geometrically meaningful way by leveraging Neural Controlled Differential Equations guided with SO(3) Savitzky-Golay paths. Existing extrapolation methods often rely on energy conservation or constant velocity assumptions, limiting their applicability in real-world scenarios involving non-conservative forces. In contrast, our approach is agnostic to energy and momentum conservation while being robust to input noise, making it applicable to complex, non-inertial systems. Our approach is easily integrated as a module in existing pipelines and generalizes well to trajectories with unknown physical parameters. By learning to approximate object dynamics from noisy states during training, our model attains robust extrapolation capabilities in simulation and various real-world settings.

We introduce ForeSight, a novel joint detection and forecasting framework for vision-based 3D perception in autonomous vehicles. Traditional approaches treat detection and forecasting as separate sequential tasks, limiting their ability to leverage temporal cues. ForeSight addresses this limitation with a multi-task streaming and bidirectional learning approach, allowing detection and forecasting to share query memory and propagate information seamlessly. The forecast-aware detection transformer enhances spatial reasoning by integrating trajectory predictions from a multiple hypothesis forecast memory queue, while the streaming forecast transformer improves temporal consistency using past forecasts and refined detections. Unlike tracking-based methods, ForeSight eliminates the need for explicit object association, reducing error propagation with a tracking-free model that efficiently scales across multi-frame sequences. Experiments on the nuScenes dataset show that ForeSight achieves state-of-the-art performance, achieving an EPA of 54.9%, surpassing previous methods by 9.3%, while also attaining the best mAP and minADE among multi-view detection and forecasting models.