CVPR 2025 Papers

Multi-task prompt learning has emerged as a promising technique for fine-tuning pre-trained Vision-Language Models (VLMs) to various downstream tasks. However, existing methods ignore challenges caused by spurious correlations and dynamic task relationships, which may reduce the model performance. To tackle these challenges, we propose JSCPT, a novel approach for Joint Scheduling of Causal Prompts and Tasks to enhance multi-task prompt learning. Specifically, we first design a Multi-Task Vison-Language Prompt (MTVLP) model, which learns task-shared and task-specific vison-language prompts and selects useful prompt features via causal intervention, alleviating spurious correlations. Then, we propose the task-prompt scheduler that models inter-task affinities and assesses the causal effect of prompt features to optimize the multi-task prompt learning process. Finally, we formulate the scheduler and the multi-task prompt learning process as a bi-level optimization problem to optimize prompts and tasks adaptively. In the lower optimization, MTVLP is updated with the scheduled gradient, while in the upper optimization, the scheduler is updated with the implicit gradient. Extensive experiments show the superiority of our proposed JSCPT approach over several baselines in terms of multi-task prompt learning for pre-trained VLMs.

Vision-centric perception systems often struggle with unpredictable and coupled weather degradations in the wild. Current solutions are often limited, as they either depend on specific degradation priors or suffer from significant domain gaps. To enable robust and autonomous operation in real-world conditions, we propose JarvisIR, a VLM-powered agent that leverages the VLM (e.g., Llava-Llama3) as a controller to manage multiple expert restoration models. To further enhance system robustness, reduce hallucinations, and improve generalizability in real-world adverse weather, JarvisIR employs a novel two-stage framework consisting of supervised fine-tuning and human feedback alignment. Specifically, to address the lack of paired data in real-world scenarios, the human feedback alignment enables the VLM to be fine-tuned effectively on large-scale real-world data in an unsupervised manner. To support the training and evaluation of JarvisIR, we introduce CleanBench, a comprehensive dataset consisting of high-quality and large-scale instruction-responses pairs, including 150K synthetic entries and 80K real entries. Extensive experiments demonstrate that JarvisIR exhibits superior decision-making and restoration capabilities. Compared with existing methods, it achieves a 50% improvement in the average of all perception metrics on CleanBench-Real.

We present JanusFlow, a powerful framework that unifies image understanding and generation in a single model.JanusFlow introduces a minimalist architecture that integrates autoregressive language models with rectified flow, a state-of-the-art method in generative modeling.Our key finding demonstrates that rectified flow can be straightforwardly trained within the large language model framework, eliminating the need for complex architectural modifications.To further improve the performance of our unified model, we adopt two key strategies: (i) decoupling the understanding and generation encoders, and (ii) aligning their representations during unified training.Extensive experiments show that JaunsFlow achieves comparable or superior performance to specialized models in their respective domains, while significantly outperforming existing unified approaches.This work represents a step toward more efficient and versatile vision-language models.

We introduce Janus, an autoregressive framework that unifies multimodal understanding and generation. Prior research often relies on a single visual encoder for both tasks, such as Chameleon. However, due to the differing levels of information granularity required by multimodal understanding and generation, this approach can lead to suboptimal performance, particularly in multimodal understanding. To address this issue, we decouple visual encoding into separate pathways, while still leveraging a single, unified transformer architecture for processing. The decoupling not only alleviates the conflict between the visual encoder's roles in understanding and generation, but also enhances the framework's flexibility. For instance, both the multimodal understanding and generation components can independently select their most suitable encoding methods. Experiments show that Janus surpasses previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus make it a strong candidate for next-generation unified multimodal models.

We propose a novel Iterative Predictor-Critic Code Decoding framework for real-world image dehazing, abbreviated as IPC-Dehaze, which leverages the high-quality codebook prior encapsulated in a pre-trained VQGAN. Apart from previous codebook-based methods that rely on one-shot decoding, our method utilizes high-quality codes obtained in the previous iteration to guide the prediction of the Code-Predictor in the subsequent iteration, improving code prediction accuracy and ensuring stable dehazing performance. Our idea stems from the observations that 1) the degradation of hazy images varies with haze density and scene depth, and 2) clear regions play crucial cues in restoring dense haze regions. However, it is nontrivial to progressively refine the obtained codes in subsequent iterations, owing to the difficulty in determining which codes should be retained or replaced at each iteration. Another key insight of our study is to propose Code-Critic to capture interrelations among codes. The Code-Critic is used to evaluate code correlations and then resample a set of codes with the highest mask scores, i.e., a higher score indicates that the code is more likely to be rejected, which helps retain more accurate codes and predict difficult ones. Extensive experiments demonstrate the superiority of our method over state-of-the-art methods in real-world dehazing. Our code will be made publicly available.

Low-rank adaptations (LoRA) are widely used to fine-tune large models across various domains for specific downstream tasks. While task-specific LoRAs are often available, concerns about data privacy and intellectual property can restrict access to training data, limiting the acquisition of a multi-task model through gradient-based training. In response, LoRA merging presents an effective solution by combining multiple LoRAs into a unified adapter while maintaining data privacy. Prior works on LoRA merging primarily frame it as an optimization problem, yet these approaches face several limitations, including the rough assumption about input features utilized in optimization, massive sample requirements, and the unbalanced optimization objective. These limitations can significantly degrade performance. To address these, we propose a novel optimization-based method, named IterIS: 1) We formulate LoRA merging as an advanced optimization problem to mitigate the rough assumption. Additionally, we employ an iterative inference-solving framework in our algorithm. It can progressively refine the optimization objective for improved performance. 2) We introduce an efficient regularization term to reduce the need for massive sample requirements (requiring only 1-5% of the unlabeled samples compared to prior methods). 3) We utilize adaptive weights in the optimization objective to mitigate potential unbalances in LoRA merging process. Our method demonstrates significant improvements over multiple baselines and state-of-the-art methods in composing tasks for text-to-image diffusion, vision-language models, and large language models. Furthermore, our layer-wise algorithm can achieve convergence with minimal steps, ensuring efficiency in both memory and computation.

The platonic representation hypothesis suggests that vision and language embeddings become more homogeneous as model and dataset sizes increase. In particular, pairwise distances within each modality become more similar. This suggests that as foundation models mature, it may become possible to match vision and language embeddings in a fully unsupervised fashion, i.e. without parallel data. We present the first feasibility study, and investigate conformity of existing vision and language foundation models in the context of unsupervised, or "blind", matching. First, we formulate unsupervised matching as a quadratic assignment problem and introduce a novel heuristic that outperforms previous solvers. We also develop a technique to find optimal matching problems, for which a non-trivial match is very likely. Second, we conduct an extensive study deploying a range of vision and language models on four datasets. Our analysis reveals that for many problem instances, vision and language representations can be indeed matched without supervision. This finding opens up the exciting possibility of embedding semantic knowledge into other modalities virtually annotation-free. As a proof of concept, we showcase an unsupervised classifier, which achieves non-trivial classification accuracy without any image-text annotation.

Multimodal large language models (MLLMs) act as essential interfaces, connecting humans with AI technologies in multimodal applications. However, current MLLMs face challenges in accurately interpreting object orientation in images due to inconsistent orientation annotations in training data, hindering the development of a coherent orientation understanding. To overcome this, we propose egocentric instruction tuning, which aligns MLLMs' orientation understanding with the user's perspective, based on a consistent annotation standard derived from the user's egocentric viewpoint. We first generate egocentric instruction data that leverages MLLMs' ability to recognize object details and applies prior knowledge for orientation understanding. Using this data, we perform instruction tuning to enhance the model's capability for accurate orientation interpretation. In addition, we introduce EgoOrientBench, a benchmark that evaluates MLLMs' orientation understanding across three tasks using images collected from diverse domains. Experimental results on this benchmark show that egocentric instruction tuning significantly improves orientation understanding without compromising overall MLLM performance. The instruction data and benchmark dataset are available on our project page at https://github.com/jhCOR/EgoOrientBench.

The current state-of-the-art video generative models can produce commercial-grade videos with highly realistic details. However, they still struggle to coherently present multiple sequential events in specific short stories, which is foreseeable an essential capability for future long video generation scenarios. For example, top T2V generative models still fail to generate a video of the short simple story "how to put an elephant into a refrigerator." While existing detail-oriented benchmarks primarily focus on fine-grained metrics like aesthetic quality and spatial-temporal consistency, they fall short of evaluating models' abilities to handle event-level story presentation. To address this gap, we introduce StoryEval, a story-oriented benchmark specifically designed to assess text-to-video (T2V) models' story-completion capabilities. StoryEval features 423 prompts spanning 7 classes, each representing short stories composed of 2-4 consecutive events. We employ Vision-Language Models, such as GPT-4o and LLaVA-OV-Chat-72B, to verify the completion of each event in the generated videos, applying a unanimous voting method to enhance reliability. Our methods ensure high alignment with human evaluations, and the evaluation of 11 models reveals its challenge, with none exceeding an average story-completion rate of 50%. StoryEval provides a new benchmark for advancing T2V models and highlights the challenges and opportunities in developing next-generation solutions for coherent story-driven video generation. Project website is available at https://ypwang61.github.io/project/StoryEval.

Data Augmentation (DA), i.e., synthesizing faithful and diverse samples to expand the original training set, is a prevalent and effective strategy to improve the performance of various data-scarce tasks. With the powerful image generation ability, diffusion-based DA has shown strong performance gains on different image classification benchmarks. In this paper, we analyze today's diffusion-based DA methods, and argue that they cannot take account of both faithfulness and diversity, which are two critical keys for generating high-quality samples and boosting classification performance. To this end, we propose a novel Diffusion-based DA method: Diff-II. Specifically, it consists of three steps: 1) Category concepts learning: Learning concept embeddings for each category. 2) Inversion interpolation: Calculating the inversion for each image, and conducting circle interpolation for two randomly sampled inversions from the same category. 3) Two-stage denoising: Using different prompts to generate synthesized images in a coarse-to-fine manner. Extensive experiments on various data-scarce image classification tasks (e.g., few-shot, long-tailed, and out-of-distribution classification) have demonstrated its effectiveness over state-of-the-art diffusion-based DA methods.

Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNets, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNets while achieving better interpretability.

Chain-of-Thought (CoT) prompting elicits large language models (LLMs) to produce a series of intermediate reasoning steps before arriving at the final answer.However, when transitioning to vision-language models (VLMs), their text-only rationales struggle to express the fine-grained associations with the original image.In this paper, we propose an image-incorporated multimodal Chain-of-Thought, named Interleaved-modal Chain-of-Thought (ICoT), which generates sequential reasoning steps consisting of paired visual and textual rationales to infer the final answer.Intuitively, the novel ICoT requires VLMs to enable the generation of fine-grained interleaved-modal content, which is hard for current VLMs to fulfill.Considering that the required visual information is usually part of the input image, we propose Attention-driven Selection (ADS) to realize ICoT over existing VLMs.ADS intelligently inserts regions of the input image to generate the interleaved-modal reasoning steps with ignorable additional latency.ADS relies solely on the attention map of VLMs without the need for parameterization, and therefore it is a plug-and-play strategy that can be generalized to a spectrum of VLMs.We apply ADS to realize ICoT on two popular VLMs of different architectures.Extensive evaluations of three benchmarks have shown that ICoT prompting achieves substantial performance (up to 14%) and interpretability improvements compared to existing multimodal CoT prompting methods.

The ability to interpret and intervene model decisions is important for the adoption of computer-aided diagnosis methods in clinical workflows. Recent concept-based methods link the model predictions with interpretable concepts and modify their activation scores to interact with the model. However, these concepts are at the image level, which hinders the model from pinpointing the exact patches the concepts are activated. Alternatively, prototype-based methods learn representations from training image patches and compare these with test image patches, using the similarity scores for final class prediction. However, interpreting the underlying concepts of these patches can be challenging and often necessitates post-hoc guesswork. To address this issue, this paper introduces the novel Concept-based Similarity Reasoning network (CSR), which offers (i) patch-level prototype with intrinsic concept interpretation, and (ii) spatial interactivity. First, the proposed CSR provides localized explanation by grounding prototypes of each concept on image regions. Second, our model introduces novel spatial-level interaction, allowing doctors to engage directly with specific image areas, making it an intuitive and transparent tool for medical imaging. CSR improves upon prior state-of-the-art interpretable methods by up to 4.% across three biomedical datasets.

We introduce InteractVLM, a novel method to estimate 3D contact points on human bodies and objects from single in-the-wild images, enabling accurate human-object joint reconstruction in 3D. This is challenging due to occlusions, depth ambiguities, and widely varying object shapes. Existing methods rely on 3D contact annotations collected via expensive motion-capture systems or tedious manual labeling, limiting scalability and generalization. To overcome this, InteractVLM harnesses the broad visual knowledge of large Vision-Language Models (VLMs), fine-tuned with limited 3D contact data. However, directly applying these models is non-trivial, as they reason only in 2D, while human-object contact is inherently 3D. Thus we introduce a novel Render-Localize-Lift module that: (1) embeds 3D body and object surfaces in 2D space via multi-view rendering, (2) trains a novel multi-view localization model (MV-Loc) to infer contacts in 2D, and (3) lifts these to 3D. Additionally, we propose a new task called Semantic Human Contact estimation, where human contact predictions are conditioned explicitly on object semantics, enabling richer interaction modeling. InteractVLM outperforms existing work on contact estimation and also facilitates 3D reconstruction from an in-the wild image.

This paper introduces a novel dataset construction pipeline that samples pairs of frames from videos and uses multimodal large language models (MLLMs) to generate editing instructions for training instruction-based image manipulation models. Video frames inherently preserve the identity of subjects and scenes, ensuring consistent content preservation during editing. Additionally, video data captures diverse, natural dynamics--such as non-rigid subject motion and complex camera movements--that are difficult to model otherwise, making it an ideal source for scalable dataset construction. Using this approach, we create a new dataset to train InstructMove, a model capable of instruction-based complex manipulations that are difficult to achieve with synthetically generated datasets. Our model demonstrates state-of-the-art performance in tasks such as adjusting subject poses, rearranging elements, and altering camera perspectives.

In this paper, we focus on the task of instruction-based image editing. Previous works like InstructPix2Pix, InstructDiffusion, and SmartEdit have explored end-to-end editing. However, two limitations still remain: First, existing datasets suffer from low resolution, poor background consistency, and overly simplistic instructions. Second, current approaches mainly condition on the text while the rich image information is underexplored, therefore inferior in complex instruction following and maintaining background consistency. Targeting these issues, we first curated the AdvancedEdit dataset using a novel data construction pipeline, formulating a large-scale dataset with high visual quality, complex instructions, and good background consistency. Then, to further inject the rich image information, we introduce a two-stream bridging mechanism utilizing both the textual and visual features reasoned by the powerful Multimodal Large Language Models (MLLM) to guide the image editing process more precisely. Extensive results demonstrate that our approach, InsightEdit, achieves state-of-the-art performance, excelling in complex instruction following and maintaining high background consistency with the original image.

Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present Insight-V, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent's generation stability and quality. Based on the popular LLaVA-NeXT model, our method shows an average improvement of 7.5% across seven challenging multi-modal benchmarks requiring visual reasoning. We also achieve a 4.2% improvement on a stronger base MLLM, highlighting the potential to further advance state-of-the-art models. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks. We will make our data and code publicly available to promote future research in this emerging field.

Leveraging Large Language Models (LLMs) for text representation has achieved significant success, but the exploration of using Multimodal LLMs (MLLMs) for multimodal representation remains limited. Previous MLLM-based representation studies have primarily focused on unifying the embedding space while neglecting the importance of multimodal alignment. As a result, their cross-modal retrieval performance falls markedly behind that of the CLIP series models. To address this, in our work, we 1) construct DeKon5M, a contrastive learning dataset enriched with dense multimodal knowledge, which efficiently enhances multimodal alignment capabilities in representation tasks. 2) design a framework for training unified representation on MLLMs. Building upon this unified representation framework and the dense knowledge dataset DeKon5M, we developed the dense knowledge representation model DeKR on Qwen2VL. Through extensive quantitative and qualitative experiments, our results demonstrate that DeKR not only aligns text, image, video, and text-image combinations within a unified embedding space but also achieves cross-modal retrieval performance comparable to SoTA CLIP series models. This fully validates the effectiveness of our approach and provides new insights for multimodal representation research.

In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques.