Chao Wen

Chao Wen, Jacqueline Staub, Adish Singla

Vision-language models (VLMs) have been explored for visual programming, where they generate code to solve visual tasks. However, most prior work focuses on visual programming for productivity; it remains unclear how well current VLMs perform on education-oriented visual programming and what factors limit their performance. To bridge this gap, we introduce TurtleAI, a benchmark containing 823 tasks curated based on real-world visual programming tasks in the Turtle Graphics domain. Solving these tasks requires models to perceive geometric patterns, reason about spatial relationships, and synthesize Python code that faithfully reproduces geometric patterns. We evaluate 20+ VLMs, including GPT-5, GPT-4o, and Qwen2-VL-72B, and find that they struggle significantly, with most achieving success rates below 30%. To address these limitations, we propose a data generation technique that requires only a small set of seed samples. Fine-tuning Qwen2-VL-72B on the resulting synthetic data yields an improvement of about 20% on real-world tasks. Our failure analysis reveals that GPT-4o struggles with spatial reasoning and precise visual replication, whereas fine-tuning primarily improves the alignment between visual reasoning and code implementation.

Chao Wen, Yinda Zhang, Chenjie Cao et al.

We study the problem of shape generation in 3D mesh representation from a small number of color images with or without camera poses. While many previous works learn to hallucinate the shape directly from priors, we adopt to further improve the shape quality by leveraging cross-view information with a graph convolution network. Instead of building a direct mapping function from images to 3D shape, our model learns to predict series of deformations to improve a coarse shape iteratively. Inspired by traditional multiple view geometry methods, our network samples nearby area around the initial mesh's vertex locations and reasons an optimal deformation using perceptual feature statistics built from multiple input images. Extensive experiments show that our model produces accurate 3D shapes that are not only visually plausible from the input perspectives, but also well aligned to arbitrary viewpoints. With the help of physically driven architecture, our model also exhibits generalization capability across different semantic categories, and the number of input images. Model analysis experiments show that our model is robust to the quality of the initial mesh and the error of camera pose, and can be combined with a differentiable renderer for test-time optimization.

Yining Zhao, Chao Wen, Zhou Xue et al.

Significant geometric structures can be compactly described by global wireframes in the estimation of 3D room layout from a single panoramic image. Based on this observation, we present an alternative approach to estimate the walls in 3D space by modeling long-range geometric patterns in a learnable Hough Transform block. We transform the image feature from a cubemap tile to the Hough space of a Manhattan world and directly map the feature to the geometric output. The convolutional layers not only learn the local gradient-like line features, but also utilize the global information to successfully predict occluded walls with a simple network structure. Unlike most previous work, the predictions are performed individually on each cubemap tile, and then assembled to get the layout estimation. Experimental results show that we achieve comparable results with recent state-of-the-art in prediction accuracy and performance. Code is available at https://github.com/Starrah/DMH-Net.

Xiaozheng Zheng, Zhuo Su, Chao Wen et al.

To bridge the physical and virtual worlds for rapidly developed VR/AR applications, the ability to realistically drive 3D full-body avatars is of great significance. Although real-time body tracking with only the head-mounted displays (HMDs) and hand controllers is heavily under-constrained, a carefully designed end-to-end neural network is of great potential to solve the problem by learning from large-scale motion data. To this end, we propose a two-stage framework that can obtain accurate and smooth full-body motions with the three tracking signals of head and hands only. Our framework explicitly models the joint-level features in the first stage and utilizes them as spatiotemporal tokens for alternating spatial and temporal transformer blocks to capture joint-level correlations in the second stage. Furthermore, we design a set of loss terms to constrain the task of a high degree of freedom, such that we can exploit the potential of our joint-level modeling. With extensive experiments on the AMASS motion dataset and real-captured data, we validate the effectiveness of our designs and show our proposed method can achieve more accurate and smooth motion compared to existing approaches.

Pengfei Ren, Chao Wen, Xiaozheng Zheng et al.

Reconstructing interacting hands from a single RGB image is a very challenging task. On the one hand, severe mutual occlusion and similar local appearance between two hands confuse the extraction of visual features, resulting in the misalignment of estimated hand meshes and the image. On the other hand, there are complex spatial relationship between interacting hands, which significantly increases the solution space of hand poses and increases the difficulty of network learning. In this paper, we propose a decoupled iterative refinement framework to achieve pixel-alignment hand reconstruction while efficiently modeling the spatial relationship between hands. Specifically, we define two feature spaces with different characteristics, namely 2D visual feature space and 3D joint feature space. First, we obtain joint-wise features from the visual feature map and utilize a graph convolution network and a transformer to perform intra- and inter-hand information interaction in the 3D joint feature space, respectively. Then, we project the joint features with global information back into the 2D visual feature space in an obfuscation-free manner and utilize the 2D convolution for pixel-wise enhancement. By performing multiple alternate enhancements in the two feature spaces, our method can achieve an accurate and robust reconstruction of interacting hands. Our method outperforms all existing two-hand reconstruction methods by a large margin on the InterHand2.6M dataset.

Chao Wen, Tung Phung, Pronita Mehrotra et al.

Generative AI has democratized content creation, but popular chatbot-based interfaces often prioritize execution, generating fully rendered artifacts right away. This issue can lead to premature convergence and design fixation, where users are being anchored to initial outputs. Recent works have proposed new interfaces to address this issue by supporting exploration, though typically constrained to be semantically close to a user's initial task framing, potentially limiting the creativity of the outcomes. We examine an approach grounded in the Geneplore model of creative cognition and instantiate it in a human-AI co-creation system, HAICo, for creative image generation. HAICo explicitly structures the creative process into two switchable modes: DIVERGENT mode scaffolds the broad exploration of remote conceptual ideas; CONVERGENT mode supports a targeted refinement of selected ideas. Through a within-subjects study (N=24) on a poster image creation task, we demonstrate that HAICo outperforms ChatGPT across multiple dimensions of creativity and usability. Our results highlight the critical need to shift from pure execution-focused chatbots to scaffolded co-creation systems that actively guide exploration and foster the creative process.

Zhongcan Li, Ping Huang, Chao Wen et al.

Railway operations involve different types of entities (stations, trains, etc.), making the existing graph/network models with homogenous nodes (i.e., the same kind of nodes) incapable of capturing the interactions between the entities. This paper aims to develop a heterogeneous graph neural network (HetGNN) model, which can address different types of nodes (i.e., heterogeneous nodes), to investigate the train delay evolution on railway networks. To this end, a graph architecture combining the HetGNN model and the GraphSAGE homogeneous GNN (HomoGNN), called SAGE-Het, is proposed. The aim is to capture the interactions between trains, trains and stations, and stations and other stations on delay evolution based on different edges. In contrast to the traditional methods that require the inputs to have constant dimensions (e.g., in rectangular or grid-like arrays) or only allow homogeneous nodes in the graph, SAGE-Het allows for flexible inputs and heterogeneous nodes. The data from two sub-networks of the China railway network are applied to test the performance and robustness of the proposed SAGE-Het model. The experimental results show that SAGE-Het exhibits better performance than the existing delay prediction methods and some advanced HetGNNs used for other prediction tasks; the predictive performances of SAGE-Het under different prediction time horizons (10/20/30 min ahead) all outperform other baseline methods; Specifically, the influences of train interactions on delay propagation are investigated based on the proposed model. The results show that train interactions become subtle when the train headways increase . This finding directly contributes to decision-making in the situation where conflict-resolution or train-canceling actions are needed.

Xiaozheng Zheng, Chao Wen, Zhou Xue et al.

Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single-view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the ``groupthink'' effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single-view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single-view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under the same settings.

Xiaozheng Zheng, Chao Wen, Zhaohu Li et al.

In this paper, we present a novel 3D head avatar creation approach capable of generalizing from few-shot in-the-wild data with high-fidelity and animatable robustness. Given the underconstrained nature of this problem, incorporating prior knowledge is essential. Therefore, we propose a framework comprising prior learning and avatar creation phases. The prior learning phase leverages 3D head priors derived from a large-scale multi-view dynamic dataset, and the avatar creation phase applies these priors for few-shot personalization. Our approach effectively captures these priors by utilizing a Gaussian Splatting-based auto-decoder network with part-based dynamic modeling. Our method employs identity-shared encoding with personalized latent codes for individual identities to learn the attributes of Gaussian primitives. During the avatar creation phase, we achieve fast head avatar personalization by leveraging inversion and fine-tuning strategies. Extensive experiments demonstrate that our model effectively exploits head priors and successfully generalizes them to few-shot personalization, achieving photo-realistic rendering quality, multi-view consistency, and stable animation.

Hangyu Lin, Chao Wen, Chengming Xu et al.

Flow matching based video generative models have been increasingly relying on prepended Vision-Language Models (VLMs) to handle complex, instruction-based video editing. The prevailing assumption underlying this paradigm is that a connector module can seamlessly align the VLM's rich multi-modal reasoning with the original text embedding space of DiTs. However, we hypothesize that this alignment acts as a severe semantic bottleneck, degrading fine-grained structural variables. Verifying this is challenging, as end-to-end evaluations conflate alignment failures with generation errors, and natural datasets lack disentangled annotations. To rigorously investigate this, we propose a controlled data processing pipeline based on video composition that results in TRACE-Edit, a diagnostic dataset focusing on relation-based editing. Leveraging this dataset, we propose a comprehensive diagnostic protocol to analyze two important designs of meta-query and connector in the existing video editing models. Systematic evaluation of four representative model cases reveals that fine-grained structural semantics can be severely degraded during alignment. Our findings overturn the assumption of lossless semantic transfer, identifying the VLM-to-DiT alignment as a major bottleneck and providing a new diagnostic foundation for future multi-modal alignment architectures.

Jiashun Wang, Chao Wen, Yanwei Fu et al.

Pose transfer has been studied for decades, in which the pose of a source mesh is applied to a target mesh. Particularly in this paper, we are interested in transferring the pose of source human mesh to deform the target human mesh, while the source and target meshes may have different identity information. Traditional studies assume that the paired source and target meshes are existed with the point-wise correspondences of user annotated landmarks/mesh points, which requires heavy labelling efforts. On the other hand, the generalization ability of deep models is limited, when the source and target meshes have different identities. To break this limitation, we proposes the first neural pose transfer model that solves the pose transfer via the latest technique for image style transfer, leveraging the newly proposed component -- spatially adaptive instance normalization. Our model does not require any correspondences between the source and target meshes. Extensive experiments show that the proposed model can effectively transfer deformation from source to target meshes, and has good generalization ability to deal with unseen identities or poses of meshes. Code is available at https://github.com/jiashunwang/Neural-Pose-Transfer .

Xiaozheng Zheng, Chao Wen, Zhuo Su et al.

In this paper, we delve into the creation of one-shot hand avatars, attaining high-fidelity and drivable hand representations swiftly from a single image. With the burgeoning domains of the digital human, the need for quick and personalized hand avatar creation has become increasingly critical. Existing techniques typically require extensive input data and may prove cumbersome or even impractical in certain scenarios. To enhance accessibility, we present a novel method OHTA (One-shot Hand avaTAr) that enables the creation of detailed hand avatars from merely one image. OHTA tackles the inherent difficulties of this data-limited problem by learning and utilizing data-driven hand priors. Specifically, we design a hand prior model initially employed for 1) learning various hand priors with available data and subsequently for 2) the inversion and fitting of the target identity with prior knowledge. OHTA demonstrates the capability to create high-fidelity hand avatars with consistent animatable quality, solely relying on a single image. Furthermore, we illustrate the versatility of OHTA through diverse applications, encompassing text-to-avatar conversion, hand editing, and identity latent space manipulation.

Muxin Zhang, Qiao Feng, Zhuo Su et al.

3D human generation is increasingly significant in various applications. However, the direct use of 2D generative methods in 3D generation often results in losing local details, while methods that reconstruct geometry from generated images struggle with global view consistency. In this work, we introduce Joint2Human, a novel method that leverages 2D diffusion models to generate detailed 3D human geometry directly, ensuring both global structure and local details. To achieve this, we employ the Fourier occupancy field (FOF) representation, enabling the direct generation of 3D shapes as preliminary results with 2D generative models. With the proposed high-frequency enhancer and the multi-view recarving strategy, our method can seamlessly integrate the details from different views into a uniform global shape. To better utilize the 3D human prior and enhance control over the generated geometry, we introduce a compact spherical embedding of 3D joints. This allows for an effective guidance of pose during the generation process. Additionally, our method can generate 3D humans guided by textual inputs. Our experimental results demonstrate the capability of our method to ensure global structure, local details, high resolution, and low computational cost simultaneously. More results and the code can be found on our project page at http://cic.tju.edu.cn/faculty/likun/projects/Joint2Human.

Yuang Wang, Chao Wen, Haoyu Guo et al.

We present visual action prompts, a unified action representation for action-to-video generation of complex high-DoF interactions while maintaining transferable visual dynamics across domains. Action-driven video generation faces a precision-generality trade-off: existing methods using text, primitive actions, or coarse masks offer generality but lack precision, while agent-centric action signals provide precision at the cost of cross-domain transferability. To balance action precision and dynamic transferability, we propose to "render" actions into precise visual prompts as domain-agnostic representations that preserve both geometric precision and cross-domain adaptability for complex actions; specifically, we choose visual skeletons for their generality and accessibility. We propose robust pipelines to construct skeletons from two interaction-rich data sources - human-object interactions (HOI) and dexterous robotic manipulation - enabling cross-domain training of action-driven generative models. By integrating visual skeletons into pretrained video generation models via lightweight fine-tuning, we enable precise action control of complex interaction while preserving the learning of cross-domain dynamics. Experiments on EgoVid, RT-1 and DROID demonstrate the effectiveness of our proposed approach. Project page: https://zju3dv.github.io/VAP/.

Chao Wen, Jacqueline Staub, Adish Singla

Large language and multimodal models have shown remarkable success on various benchmarks focused on specific skills such as general-purpose programming, math word problem-solving, and visual question answering. However, it is unclear how well these models perform on tasks that require a combination of these skills. In this paper, we curate a novel program synthesis benchmark based on the real-world tasks in the XLogoOnline visual programming environment. Each task requires a combination of different skills such as spatial planning, basic programming, and logical reasoning. Our evaluation shows that current state-of-the-art models like GPT-4V and Llama3-70B struggle to solve these tasks, achieving only 20% and 2.35% success rates, respectively. Next, we develop a fine-tuning pipeline to boost the performance of models by leveraging a large-scale synthetic training dataset with over 80,000 tasks. Moreover, we showcase how emulator-driven feedback can be used to design a curriculum over training data distribution, through which a fine-tuned Llama3-8B drastically outperforms GPT-4V and Llama3-70B models. Finally, we provide an in-depth failure analysis to understand the limitations of different models. We will publicly release the benchmark for future research on program synthesis in visual programming.

Chao Wen, Miao Xu, Zhilin Zhang et al.

In online advertising, auto-bidding has become an essential tool for advertisers to optimize their preferred ad performance metrics by simply expressing high-level campaign objectives and constraints. Previous works designed auto-bidding tools from the view of single-agent, without modeling the mutual influence between agents. In this paper, we instead consider this problem from a distributed multi-agent perspective, and propose a general $\underline{M}$ulti-$\underline{A}$gent reinforcement learning framework for $\underline{A}$uto-$\underline{B}$idding, namely MAAB, to learn the auto-bidding strategies. First, we investigate the competition and cooperation relation among auto-bidding agents, and propose a temperature-regularized credit assignment to establish a mixed cooperative-competitive paradigm. By carefully making a competition and cooperation trade-off among agents, we can reach an equilibrium state that guarantees not only individual advertiser's utility but also the system performance (i.e., social welfare). Second, to avoid the potential collusion behaviors of bidding low prices underlying the cooperation, we further propose bar agents to set a personalized bidding bar for each agent, and then alleviate the revenue degradation due to the cooperation. Third, to deploy MAAB in the large-scale advertising system with millions of advertisers, we propose a mean-field approach. By grouping advertisers with the same objective as a mean auto-bidding agent, the interactions among the large-scale advertisers are greatly simplified, making it practical to train MAAB efficiently. Extensive experiments on the offline industrial dataset and Alibaba advertising platform demonstrate that our approach outperforms several baseline methods in terms of social welfare and revenue.

Tesfaye Hailemariam Yimer, Chao Wen, Xiaozhuo Yu et al.

When driving,it is vital to maintain the right following distance between the vehicles to avoid rear-end collisions. The minimum safe distance depends on many factors, however, in this study the safe distance between the human-driven vehicles and a fully autonomous vehicle at a sudden stop by an automatic emergency brake was studied based on the human driver ability to react in an accident, the vehicles' braking system performance, and the speed of vehicles. For this approach, a safe distance car-following model was proposed to describe the safe distance between vehicles on a single lane dry road under conditions where both vehicles keep moving at a constant speed, and a lead autonomous vehicle suddenly stops by automatic emergency braking at an imminent incident. The proposed model then finally was being tested using MATLAB simulation, and results showed that confirmed the effectiveness of this model and the influence of driving speed and inter-vehicle distance on the rear-end collision was also indicated as well compared with the two and three seconds rule of safe following distance. The three seconds safe distance following rules is safe to be applied for all speed limits; however, the two seconds can be used on speed limits up to 45 Km/hr. A noticeable increase in rear-end collision was observed according to the simulation results if a car follows a driverless vehicle with two seconds rule above 45 km/hr.

Xinghu Yao, Chao Wen, Yuhui Wang et al.

Learning a stable and generalizable centralized value function (CVF) is a crucial but challenging task in multi-agent reinforcement learning (MARL), as it has to deal with the issue that the joint action space increases exponentially with the number of agents in such scenarios. This paper proposes an approach, named SMIX($λ$), to address the issue using an efficient off-policy centralized training method within a flexible learner search space. As importance sampling for such off-policy training is both computationally costly and numerically unstable, we proposed to use the $λ$-return as a proxy to compute the TD error. With this new loss function objective, we adopt a modified QMIX network structure as the base to train our model. By further connecting it with the ${Q(λ)}$ approach from an unified expectation correction viewpoint, we show that the proposed SMIX($λ$) is equivalent to ${Q(λ)}$ and hence shares its convergence properties, while without being suffered from the aforementioned curse of dimensionality problem inherent in MARL. Experiments on the StarCraft Multi-Agent Challenge (SMAC) benchmark demonstrate that our approach not only outperforms several state-of-the-art MARL methods by a large margin, but also can be used as a general tool to improve the overall performance of other CTDE-type algorithms by enhancing their CVFs.

Chao Wen, Yinda Zhang, Zhuwen Li et al.

We study the problem of shape generation in 3D mesh representation from a few color images with known camera poses. While many previous works learn to hallucinate the shape directly from priors, we resort to further improving the shape quality by leveraging cross-view information with a graph convolutional network. Instead of building a direct mapping function from images to 3D shape, our model learns to predict series of deformations to improve a coarse shape iteratively. Inspired by traditional multiple view geometry methods, our network samples nearby area around the initial mesh's vertex locations and reasons an optimal deformation using perceptual feature statistics built from multiple input images. Extensive experiments show that our model produces accurate 3D shape that are not only visually plausible from the input perspectives, but also well aligned to arbitrary viewpoints. With the help of physically driven architecture, our model also exhibits generalization capability across different semantic categories, number of input images, and quality of mesh initialization.

Yuhui Wang, Hao He, Chao Wen et al.

Proximal policy optimization (PPO) is one of the most successful deep reinforcement-learning methods, achieving state-of-the-art performance across a wide range of challenging tasks. However, its optimization behavior is still far from being fully understood. In this paper, we show that PPO could neither strictly restrict the likelihood ratio as it attempts to do nor enforce a well-defined trust region constraint, which means that it may still suffer from the risk of performance instability. To address this issue, we present an enhanced PPO method, named Truly PPO. Two critical improvements are made in our method: 1) it adopts a new clipping function to support a rollback behavior to restrict the difference between the new policy and the old one; 2) the triggering condition for clipping is replaced with a trust region-based one, such that optimizing the resulted surrogate objective function provides guaranteed monotonic improvement of the ultimate policy performance. It seems, by adhering more truly to making the algorithm proximal - confining the policy within the trust region, the new algorithm improves the original PPO on both sample efficiency and performance.