CVPR 2025 论文列表

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.

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.

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

Image tiling--the seamless connection of disparate images to create a coherent visual field--is crucial for applications such as texture creation, video game asset development, and digital art. Traditionally, tiles have been constructed manually, a method that poses significant limitations in scalability and flexibility. Recent research has attempted to automate this process using generative models. However, current approaches primarily focus on tiling textures and manipulating models for single-image generation, without inherently supporting the creation of multiple interconnected tiles across diverse domains.This paper presents Tiled Diffusion, a novel approach that extends the capabilities of diffusion models to accommodate the generation of cohesive tiling patterns across various domains of image synthesis that require tiling. Our method supports a wide range of tiling scenarios, from self-tiling to complex many-to-many connections, enabling seamless integration of multiple images.Tiled Diffusion automates the tiling process, eliminating the need for manual intervention and enhancing creative possibilities in various applications, such as seamlessly tiling of existing images, tiled texture creation, and 360deg synthesis.

We consider the task of Image-to-Video (I2V) generation, which involves transforming static images into realistic video sequences based on a textual description. While recent advancements produce photorealistic outputs, they frequently struggle to create videos with accurate and consistent object motion, especially in multi-object scenarios. To address these limitations, we propose a two-stage compositional framework that decomposes I2V generation into: (i) An explicit intermediate representation generation stage, followed by (ii) A video generation stage that is conditioned on this representation. Our key innovation is the introduction of a mask-based motion trajectory as an intermediate representation, that captures both semantic object information and motion, enabling an expressive but compact representation of motion and semantics. To incorporate the learned representation in the second stage, we utilize object-level attention objectives. Specifically, we consider a spatial, per-object, masked-cross attention objective, integrating object-specific prompts into corresponding latent space regions and a masked spatio-temporal self-attention objective, ensuring frame-to-frame consistency for each object. We evaluate our method on challenging benchmarks with multi-object and high-motion scenarios and empirically demonstrate that the proposed method achieves state-of-the-art results in temporal coherence, motion realism, and text-prompt faithfulness. Additionally, we introduce SA-V-128, a new challenging benchmark for single-object and multi-object I2V generation, and demonstrate our method's superiority on this benchmark. Project page is available at https://guyyariv.github.io/TTM/

3D reconstruction of highly deformable surfaces (e.g. cloths) from monocular RGB videos is a challenging problem, and no solution provides a consistent and accurate recovery of fine-grained surface details. To account for the ill-posed nature of the setting, existing methods use deformation models with statistical, neural, or physical priors. They also predominantly rely on nonadaptive discrete surface representations (e.g. polygonal meshes), perform frame-by-frame optimisation leading to error propagation, and suffer from poor gradients of the mesh-based differentiable renderers. Consequently, fine surface details such as cloth wrinkles are often not recovered with the desired accuracy. In response to these limitations, we propose Thin-Shell-SfT, a new method for non-rigid 3D tracking that represents a surface as an implicit and continuous spatiotemporal neural field. We incorporate continuous thin shell physics prior based on the Kirchhoff-Love model for spatial regularisation, which starkly contrasts the discretised alternatives of earlier works. Lastly, we leverage 3D Gaussian splatting to differentiably render the surface into image space and optimise the deformations based on analysis-by-synthesis principles. Our Thin-Shell-SfT outperforms prior works qualitatively and quantitatively thanks to our continuous surface formulation in conjunction with a specially tailored simulation prior and surface-induced 3D Gaussians. See our project page at https://4dqv.mpi-inf.mpg.de/ThinShellSfT.

We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. It represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms an automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing.

Single-image super-resolution (SISR) remains challenging due to the inherent difficulty of recovering fine-grained details and preserving perceptual quality from low-resolution inputs. Existing methods often rely on limited image priors, leading to suboptimal results. We propose a novel approach that leverages the rich contextual information available in multiple modalities -- including depth, segmentation, edges, and text prompts -- to learn a powerful generative prior for SISR within a diffusion model framework. We introduce a flexible network architecture that effectively fuses multimodal information, accommodating an arbitrary number of input modalities without requiring significant modifications to the diffusion process. Crucially, we mitigate hallucinations, often introduced by text prompts, by using spatial information from other modalities to guide regional text-based conditioning. Each modality's guidance strength can also be controlled independently, allowing steering outputs toward different directions, such as increasing bokeh through depth or adjusting object prominence via segmentation. Extensive experiments demonstrate that our model surpasses state-of-the-art generative SISR methods, achieving superior visual quality and fidelity.

In image segmentation and specifically in medical image segmentation, the soft-Dice loss is often chosen instead of the more traditional cross-entropy loss to improve performance with respect to the Dice metric.Experimental work supporting this claim exists, but how and why the two loss functions lead to different predictions is not well understood.This paper explains the observed discrepancy as a consequence of how those loss functions are effected by label noise and what threshold is used to convert the predicted soft segmentation to predicted labels.In particular, it is shown (i) how the optimal solutions to the two loss functions diverge as the noise is increased, and (ii) how the optimal solutions to soft-Dice can be recovered by thresholding the solutions to cross-entropy with an a priori unknown but efficiently computable threshold.The theoretical results are supported by numerical experiments and it is concluded that cross-entropy with the alternative threshold yields the stability and informative label probability maps associated with cross-entropy without sacrificing the performance of soft-Dice.

Text-to-image models such as Stable Diffusion, DALL*E, and Midjourney have gained immense popularity lately. However, they are trained on vast amounts of data that may include private, explicit, or copyrighted material used without permission, raising serious legal and ethical concerns. In light of the recent regulations aimed at protecting individual data privacy, there has been a surge in Machine Unlearning methods designed to remove specific concepts from these models. However, we identify a critical flaw in these unlearning techniques: unlearned concepts will revive when the models are fine-tuned, even with general or unrelated prompts. In this paper, for the first time, through an extensive study, we demonstrate the unstable nature of existing unlearning methods in text-to-image diffusion models. We introduce a framework that includes a couple of measures for analyzing the stability of existing unlearning methods. Further, the paper offers preliminary insights into the plausible explanation for the instability of the mapping-based unlearning methods that can guide future research toward more robust unlearning techniques. Codes for implementing the proposed framework are provided.

The evolution of Text-to-video (T2V) generative models, trained on large-scale datasets, has been marked by significant progress. However, the sensitivity of T2V generative models to input prompts highlights the critical role of prompt design in influencing generative outcomes. Prior research has predominantly relied on Large Language Models (LLMs) to align user-provided prompts with the distribution of training prompts, albeit without tailored guidance encompassing prompt vocabulary and sentence structure nuances. To this end, we introduce RAPO, a novel Retrieval-Augmented Prompt Optimization framework. In order to address potential inaccuracies and ambiguous details generated by LLM-generated prompts. RAPO refines the naive prompts through dual optimization branches, selecting the superior prompt for T2V generation. The first branch augments user prompts with diverse modifiers extracted from a learned relational graph, refining them to align with the format of training prompts via a fine-tuned LLM. Conversely, the second branch rewrites the naive prompt using a pre-trained LLM following a well-defined instruction set. Extensive experiments demonstrate that RAPO can effectively enhance both the static and dynamic dimensions of generated videos, demonstrating the significance of prompt optimization for user-provided prompts. Project website: \href https://whynothaha.github.io/Prompt_optimizer/RAPO.html GitHub .

Fashion image editing is a valuable tool for designers to convey their creative ideas by visualizing design concepts. With the recent advances in text editing methods, significant progress has been made in fashion image editing. However, they face two key challenges: spurious correlations in training data often induce changes in other areas when editing an area representing the intended editing concept, and these models typically lack the ability to edit multiple concepts simultaneously. To address the above challenges, we propose a novel \underline T ext-driven \underline F ashion \underline I mage edi\underline T ing framework called T-FIT to mitigate the impact of spurious correlation by integrating counterfactual reasoning with compositional concept learning to precisely ensure compositional multi-concept fashion image editing relying solely on text descriptions. Specifically, T-FIT includes three key components. (i) Counterfactual abduction module, which learns an exogenous variable of the source image by a denoising U-Net model. (ii) Concept learning module, which identifies concepts in fashion image editing--such as clothing types and colors and projects a target concept into the space spanned from a series of textual prompts. (iii) Concept composition module, which enables simultaneous adjustments of multiple concepts by aggregating each concept's direction vector obtained from the concept learning module. Extensive experiments show that our method can achieve state-of-the-art performance on various fashion image editing tasks, including single-concept editing (e.g., sleeve length, clothing type) and multi-concept editing (e.g., color & sleeve length).

In text-to-image diffusion models, the cross-attention map of each text token indicates the specific image regions attended. Comparing these maps of syntactically related tokens provides insights into how well the generated image reflects the text prompt. For example, in the prompt, "a black car and a white clock", the cross-attention maps for "black" and "car" should focus on overlapping regions to depict a black car, while "car" and "clock" should not. Incorrect overlapping in the maps generally produces generation flaws such as missing objects and incorrect attribute binding. Our study makes the key observations investigating this issue in the existing text-to-image models: (1) the similarity in text embeddings between different tokens---used as conditioning inputs---can cause their cross-attention maps to focus on the same image regions; and (2) text embeddings often fail to faithfully capture syntactic relations already within text attention maps. As a result, such syntactic relationships can be overlooked in cross-attention module, leading to inaccurate image generation. To address this, we propose a method that directly transfers syntactic relations from the text attention maps to the cross-attention module via a test-time optimization. Our approach leverages this inherent yet unexploited information within text attention maps to enhance image-text semantic alignment across diverse prompts, without relying on external guidance. Our project page and code are available at: https://t-sam-diffusion.github.io

This paper introduces TexGarment, an efficient method for synthesizing high-quality, 3D-consistent garment textures in UV space. Traditional approaches based on 2D-to-3D mapping often suffer from 3D inconsistency, while methods learning from limited 3D data lack sufficient texture diversity. These limitations are particularly problematic in garment texture generation, where high demands exist for both detail and variety. To address these challenges, TexGarment leverages a pre-trained text-to-image diffusion Transformer model with robust generalization capabilities, introducing structural information to guide the model in generating 3D-consistent garment textures in a single inference step. Specifically, We utilize the 2D UV position map to guide the layout during the UV texture generation process, ensuring a coherent texture arrangement and enhancing it by integrating global 3D structural information from the mesh surface point cloud. This combined guidance effectively aligns 3D structural integrity with 2D layout. Our method efficiently generates high-quality, diverse UV textures in a single inference step while maintaining 3D consistency. Experimental results validate the effectiveness of TexGarment, achieving state-of-the-art performance in 3D garment texture generation.

With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone's model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/.

Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TDFusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks. The code is available at https://github.com/HaowenBai/TDFusion.

Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class--without access to pre-trained data--while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.

Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images).

Pre-trained text-to-image diffusion models are increasingly applied to real-world image super-resolution (Real-ISR) task. Given the iterative refinement nature of diffusion models, most existing approaches are computationally expensive. While methods such as SinSR and OSEDiff have emerged to condense inference steps via distillation, their performance in image restoration or details recovery is not satisfied. To address this, we propose TSD-SR, a novel distillation framework specifically designed for real-world image super-resolution, aiming to construct an efficient and effective one-step model. We first introduce the Target Score Distillation, which leverages the priors of diffusion models and real image references to achieve more realistic image restoration. Secondly, we propose a Distribution-Aware Sampling Module to make detail-oriented gradients more readily accessible, addressing the challenge of recovering fine details. Extensive experiments demonstrate that our TSD-SR has superior restoration results (most of the metrics perform the best) and the fastest inference speed (e.g. 40 times faster than SeeSR) compared to the past Real-ISR approaches based on pre-trained diffusion priors.