CVPR 2025 Papers

While MLLMs have demonstrated adequate image understanding capabilities, they still struggle with pixel-level comprehension, limiting their practical applications. Current evaluation tasks like VQA and visual grounding remain too coarse to assess fine-grained pixel comprehension accurately. Though segmentation is foundational for pixel-level understanding, existing methods often require MLLMs to generate implicit tokens, decoded through external pixel decoders. This approach disrupts the MLLM's text output space, potentially compromising language capabilities and reducing flexibility and extensibility, while failing to reflect the model's intrinsic pixel-level understanding.Thus, We introduce the Human-Like Mask Annotation Task (HLMAT), a new paradigm where MLLMs mimic human annotators using interactive segmentation tools. Modeling segmentation as a multi-step Markov Decision Process, HLMAT enables MLLMs to iteratively generate text-based click points, achieving high-quality masks without architectural changes or implicit tokens. Through this setup, we develop SegAgent, a model fine-tuned on human-like annotation trajectories, which achieves performance comparable to SOTA methods and supports additional tasks like mask refinement and annotation filtering.HLMAT provides a protocol for assessing fine-grained pixel understanding in MLLMs and introduces a vision-centric, multi-step decision-making task that facilitates exploration of MLLMs' visual reasoning abilities. Our adaptations of policy improvement method StaR and PRM guided tree search further enhance model robustness in complex segmentation tasks, laying a foundation for future advancements in fine-grained visual perception and multi-step decision-making for MLLMs.

Diffusion and flow matching models have achieved remarkable success in text-to-image generation. However, these models typically rely on the predetermined denoising schedules for all prompts. The multi-step reverse diffusion process can be regarded as a kind of chain-of-thought for generating high-quality images step by step. Therefore, diffusion models should reason for each instance to determine the optimal noise schedule adaptively, achieving high generation quality with sampling efficiency. In this paper, we introduce the Time Prediction Diffusion Model (TPDM) for this. TPDM employs a plug-and-play Time Prediction Module (TPM) that predicts the next noise level based on current latent features at each denoising step. We train the TPM using reinforcement learning to maximize a reward that encourages high final image quality while penalizing excessive denoising steps. With such an adaptive scheduler, TPDM not only generates high-quality images that are aligned closely with human preferences but also adjusts diffusion time and the number of denoising steps on the fly, enhancing both performance and efficiency. With Stable Diffusion 3 Medium architecture, TPDM achieves an aesthetic score of 5.44 and a human preference score (HPS) of 29.59, while using around 50% fewer denoising steps to achieve better performance.

We introduce Scenario Dreamer, a fully data-driven generative simulator for autonomous vehicle planning that generates both the initial traffic scene--comprising a lane graph and agent bounding boxes--and closed-loop agent behaviours. Existing methods for generating driving simulation environments encode the initial traffic scene as a rasterized image and, as such, require parameter-heavy networks that perform unnecessary computation due to many empty pixels in the rasterized scene. Moreover, we find that existing methods that employ rule-based agent behaviours lack diversity and realism. Scenario Dreamer instead employs a novel vectorized latent diffusion model for initial scene generation that directly operates on the vectorized scene elements and an autoregressive Transformer for data-driven agent behaviour simulation. Scenario Dreamer additionally supports scene extrapolation via diffusion inpainting, enabling the generation of unbounded simulation environments. Extensive experiments show that Scenario Dreamer outperforms existing generative simulators in realism and efficiency: the vectorized scene-generation base model achieves superior generation quality with around 2x fewer parameters, 6x lower generation latency, and 10x fewer GPU training hours compared to the strongest baseline. We confirm its practical utility by showing that reinforcement learning planning agents are more challenged in Scenario Dreamer environments than traditional non-generative simulation environments, especially on long and adversarial driving environments.

Recent research has shown that fine-tuning diffusion models (DMs) with arbitrary rewards, including non-differentiable ones, is feasible with reinforcement learning (RL) techniques, enabling flexible model alignment. However, applying existing RL methods to step-distilled DMs is challenging for ultra-fast (<=2-step) image generation. Our analysis suggests several limitations of policy-based RL methods such as PPO or DPO toward this goal. Based on the insights, we propose fine-tuning DMs with learned differentiable surrogate rewards. Our method, named LaSRO, learns surrogate reward models in the latent space of SDXL to convert arbitrary rewards into differentiable ones for effective reward gradient guidance. LaSRO leverages pre-trained latent DMs for reward modeling and tailors reward optimization for <=2-step image generation with efficient off-policy exploration. LaSRO is effective and stable for improving ultra-fast image generation with different reward objectives, outperforming popular RL methods including DDPO and Diffusion-DPO. We further show LaSRO's connection to value-based RL, providing theoretical insights. See our webpage at https://sites.google.com/view/lasro.

Multi-task learning (MTL) can leverage shared knowledge across tasks to improve data efficiency and generalization performance, and has been applied in various scenarios. However, task imbalance remains a major challenge for existing MTL methods. While the prior works have attempted to mitigate inter-task unfairness through loss-based and gradient-based strategies, they still exhibit imbalanced performance across tasks on common benchmarks.This key observation motivates us to consider performance-level information as an explicit fairness indicator, which can precisely reflect the current optimization status of each task, and accordingly help to adjust the gradient aggregation process.Specifically, we utilize the performance variance among tasks as the fairness indicator and introduce a dynamic weighting strategy to gradually reduce the performance variance. Based on this, we propose PIVRG, a novel performance-informed variance reduction gradient aggregation approach.Extensive experiments show that PIVRG achieves SOTA performance across various benchmarks, spanning both supervised learning and reinforcement learning tasks with task numbers ranging from 2 to 40. Results from the ablation study also show that our approach can be integrated into existing methods, significantly enhancing their performance while reducing the performance variance among tasks, thus achieving fairer optimization.

We describe a system to remove real-world reflections from images for consumer photography. Our system operates on linear (RAW) photos, and accepts an optional contextual photo looking in the opposite direction (e.g., the "selfie" camera on a mobile device). This optional photo disambiguates what should be considered the reflection. The system is trained solely on synthetic mixtures of real RAW photos, which we combine using a reflection simulation that is photometrically and geometrically accurate. Our system comprises a base model that accepts the captured photo and optional context photo as input, and runs at 256p, followed by an up-sampling model that transforms 256p images to full resolution. The system produces preview images at 1K in 4.5-6.5s on a MacBook or iPhone 14 Pro. We show SOTA results on RAW photos that were captured in the field to embody typical consumer photos, and show that training on RAW simulation data improves performance more than the architectural variations among prior works.

Recent years have witnessed the great success of denoising diffusion samplers in improving the generative capability and sampling efficiency given a pre-trained diffusion model. However, most sampling schedulers in diffusion models lack the sampling dynamics and planning capability for future generation results, leading to suboptimal solutions. To overcome this, we propose the Reinforced Active Sampling Scheduler, termed RaSS, intending to find the optimal sampling trajectory by actively planning and adjusting the sampling steps for each sampling process in time. Concretely, RaSS divides the whole sampling process into five stages and introduces a reinforcement learning (RL) agent to continuously monitor the generated instance and perceive the potential generation results, thereby achieving optimal instance- and state-adaptive sampling steps decision. Meanwhile, a sampling reward is designed to assist the planning capability of the RL agent by balancing the sampling efficiency and generation quality. The RaSS is a plug-and-play module, which is applicable to multiple denoising diffusion samplers of diffusion models. Extensive experiments on different benchmarks have shown that our RaSS can consistently improve the generation quality and efficiency across various tasks, without introducing significant computational overhead.

Video encoders optimize compression for human perception by minimizing reconstruction error under bit-rate constraints. In many modern applications such as autonomous driving, an overwhelming majority of videos serve as input for AI systems performing tasks like object recognition or segmentation, rather than being watched by humans. It is therefore useful to optimize the encoder for a downstream task instead of for perceptual image quality. However, a major challenge is how to combine such downstream optimization with existing standard video encoders, which are highly efficient and popular. Here, we address this challenge by controlling the Quantization Parameters (QPs) at the macro-block level to optimize the downstream task. This granular control allows us to prioritize encoding for task-relevant regions within each frame. We formulate this optimization problem as a Reinforcement Learning (RL) task, where the agent learns to balance long-term implications of choosing QPs on both task performance and bit-rate constraints. Notably, our policy does not require the downstream task as an input during inference, making it suitable for streaming applications and edge devices such as vehicles. We demonstrate significant improvements in two tasks, car detection, and ROI (saliency) encoding. Our approach improves task performance for a given bit rate compared to traditional task agnostic encoding methods, paving the way for more efficient task-aware video compression.

In-context learning (ICL), a predominant trend in instruction learning, aims at enhancing the performance of large language models by providing clear task guidance and examples, improving their capability in task understanding and execution. This paper investigates ICL on Large Vision-Language Models (LVLMs) and explores the policies of multi-modal demonstration selection. Existing research efforts in ICL face significant challenges: First, they rely on pre-defined demonstrations or heuristic selecting strategies based on human intuition, which are usually inadequate for covering diverse task requirements, leading to sub-optimal solutions; Second, individually selecting each demonstration fails in modeling the interactions between them, resulting in information redundancy. Unlike these prevailing efforts, we propose a new exploration-exploitation reinforcement learning framework, which explores policies to fuse multi-modal information and adaptively select adequate demonstrations as an integrated whole. The framework allows LVLMs to optimize themselves by continually refining their demonstrations through self-exploration, enabling the ability to autonomously identify and generate the most effective selection policies for in-context learning. Experimental results verify that our approach achieves significant performance improvements on four Visual Question-Answering (VQA) datasets, demonstrating its effectiveness in enhancing the generalization capability of few-shot LVLMs.

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

Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations or motion capture data. Our code is publicly available at prosepose.github.io.

Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the class-agnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel pplug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM's downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO's generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plug-and-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO.

Recovering 3D Human-Object Interaction (HOI) from single color images is challenging due to depth ambiguities, occlusions, and the huge variation in object shape and appearance. Thus, past work requires controlled settings such as known object shapes and contacts, and tackles only limited object classes. Instead, we need methods that generalize to natural images and novel object classes. We tackle this in two main ways:(1) We collect PICO-db, a new dataset of natural images uniquely paired with dense 3D contact correspondences on both body and object meshes. To this end, we use images from the recent DAMON dataset that are paired with annotated contacts, but only on a canonical 3D body. In contrast, we seek contact labels on both the body and the object. To infer these given an image, we retrieve an appropriate 3D object mesh from a database by leveraging vision foundation models. Then, we project DAMON's body contact patches onto the object via a novel method needing only 2 clicks per patch. This minimal human input establishes rich contact correspondences between bodies and objects. (2) We exploit our new dataset in a novel render-and-compare fitting method, called PICO-fit, to recover 3D body and object meshes in interaction. PICO-fit infers contact for the SMPL-X body, retrieves a likely 3D object mesh and contact from PICO-db for that object, and uses the contact to iteratively fit the 3D body and object meshes to image evidence via optimization. Uniquely, PICO-fit works well for many object categories that no existing method can tackle. This is crucial for scaling HOI understanding in the wild. Our data and code are available at https://pico.is.tue.mpg.de.

An embodied system must not only model the patterns of the external world but also understand its own motion dynamics. A motion dynamic model is essential for efficient skill acquisition and effective planning. In this work, we introduce the neural motion simulator (MoSim), a world model that predicts the future physical state of an embodied system based on current observations and actions. MoSim achieves state-of-the-art performance in physical state prediction and provides competitive performance across a range of downstream tasks. This works shows that when a world model is accurate enough and performs precise long-horizon predictions, it can facilitate efficient skill acquisition in imagined worlds and even enable zero-shot reinforcement learning. Furthermore, MoSim can transform any model-free reinforcement learning (RL) algorithm into a model-based approach, effectively decoupling physical environment modeling from RL algorithm development. This separation allows for independent advancements in RL algorithms and world modeling, significantly improving sample efficiency and enhancing generalization capabilities. Our findings highlight that world models for motion dynamics is a promising direction for developing more versatile and capable embodied systems.

We present MultiMorph, a fast and efficient method for constructing anatomical atlases on the fly. Atlases capture the canonical structure of a collection of images and are essential for quantifying anatomical variability across populations. However, current atlas construction methods often require days to weeks of computation, thereby discouraging rapid experimentation. As a result, many scientific studies rely on suboptimal, precomputed atlases from mismatched populations, negatively impacting downstream analyses. MultiMorph addresses these challenges with a feedforward model that rapidly produces high-quality, population-specific atlases in a single forward pass for any 3D brain dataset, without any fine-tuning or optimization. MultiMorph is based on a linear group-interaction layer that aggregates and shares features within the group of input images. Further, by leveraging auxiliary synthetic data, MultiMorph generalizes to new imaging modalities and population groups at test-time. Experimentally, MultiMorph outperforms state-of-the-art optimization-based and learning-based atlas construction methods in both small and large population settings, with a 100-fold reduction in time. This makes MultiMorph an accessible framework for biomedical researchers without machine learning expertise, enabling rapid, high-quality atlas generation for diverse studies.

Human hands play a central role in interacting, motivating increasing research in dexterous robotic manipulation. Data-driven embodied AI algorithms demand precise, large-scale, human-like manipulation sequences, which are challenging to obtain with conventional reinforcement learning or real-world teleoperation. To address this, we introduce ManipTrans, a novel two-stage method for efficiently transferring human bimanual skills to dexterous robotic hands in simulation. ManipTrans first pre-trains a generalist trajectory imitator to mimic hand motion, then fine-tunes a specific residual module under interaction constraints, enabling efficient learning and accurate execution of complex bimanual tasks. Experiments show that ManipTrans surpasses state-of-the-art methods in success rate, fidelity, and efficiency. Leveraging ManipTrans, we transfer multiple hand-object datasets to robotic hands, creating DexManipNet, a large-scale dataset featuring previously unexplored tasks like pen capping and bottle unscrewing. DexManipNet comprises 3.3K episodes of robotic manipulation and is easily extensible, facilitating further policy training for dexterous hands and enabling real-world deployments.

Stereo-based category-level shape and 6D pose estimation methods have the potential to generalize to a wider range of materials than RGBD methods, which often suffer from depth measurement errors. However, without explicit depth from two views, parameters to be estimated can become inherently entangled, negatively impacting performance. To address this, we propose a method that leverages global stereo consistency to constrain optimization directions and mitigate parameter entanglement. We first estimate an intra-category occupancy field to represent a unified shape across views, ensuring consistency and preventing shape ambiguity. Through a divide-and-conquer approach within global shape fitting, we fit this shape to stereo images to obtain the pose, iteratively rendering normalized depth maps and exchanging information across views. This approach improves convergence toward the correct pose and scale. We validated our method on both depth-friendly and depth-challenging materials using our S-RGBD dataset and the TOD benchmark. Our method surpasses RGBD methods on challenging objects and performs comparably on depth-friendly ones. Ablation studies confirm the effectiveness of each component.

Hiking on complex trails demands balance, agility, and adaptive decision-making over unpredictable terrain. Current humanoid research remains fragmented and inadequate for hiking: locomotion focuses on motor skills without long-term goals or situational awareness, while semantic navigation overlooks real-world embodiment and local terrain variability. We propose training humanoids to hike on complex trails, fostering integrative skill development across visual perception, decision making, and motor execution. We develop LEGO-H, a learning framework that enables a humanoid with vision to hike complex trails independently. It has two key innovations. (1) A Temporal Vision Transformer anticipates future steps to guide locomotion, unifying local movement and goal-directed navigation. (2) Latent representations of joint movement patterns combined with hierarchical metric learning allow smooth policy transfer from privileged training to real-world training. These techniques enable LEGO-H to handle diverse physical and environmental challenges without relying on predefined motion patterns. Experiments on diverse simulated hiking trails and humanoids with different morphologies demonstrate LEGO-H's robustness and versatility, establishing a strong foundation for future humanoid development.

Recent advances in text-to-image diffusion models have demonstrated impressive image generation capabilities. However, it remains challenging to control the generation process with desired properties (e.g., aesthetic quality, user intention), which can be expressed as black-box reward functions. Recent advances in text-to-image diffusion models have achieved impressive image generation capabilities. However, it remains challenging to control the generation process with desired properties (e.g., aesthetic quality, user intention), which can be expressed as black-box reward functions. In this paper, we focus on prompt adaptation, which refines the original prompt into model-preferred prompts to generate desired images. While prior work uses reinforcement learning (RL) to optimize prompts, we observe that applying RL often results in generating similar postfixes and deterministic behaviors. To this end, we introduce Prompt Adaptation with GFlowNets (PAG), a novel approach that frames prompt adaptation as a probabilistic inference problem. Our key insight is that leveraging Generative Flow Networks (GFlowNets) allows us to shift from reward maximization to sampling from an unnormalized density function, enabling both high-quality and diverse prompt generation. However, we identify that a naive application of GFlowNets suffers from mode collapse and uncovers a previously overlooked phenomenon: the progressive loss of neural plasticity in the model, which is compounded by inefficient credit assignment in sequential prompt generation. To address this critical challenge, we develop a systematic approach in PAG with flow reactivation, reward-prioritized sampling, and reward decomposition for prompt adaptation. Extensive experiments validate that PAG successfully learns to sample effective and diverse prompts for text-to-image generation. We also show that PAG exhibits strong robustness across various reward functions and transferability to different text-to-image models.