arXiv subjects

Xinggang Wang

Publications and source records attributed to Xinggang Wang.

At least 73 records · Page 4Linked to original sources

arXiv2025

RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning

Existing end-to-end autonomous driving (AD) algorithms typically follow the Imitation Learning (IL) paradigm, which faces challenges such as causal confusion and an open-loop gap. In this work, we propose RAD, a 3DGS-based closed-loop Reinforcement Learning (RL) framework for end-to-end Autonomous Driving. By leveraging 3DGS techniques, we construct a photorealistic digital replica of the real physical world, enabling the AD policy to extensively explore the state space and learn to handle out-of-distribution scenarios through large-scale trial and error. To enhance safety, we design specialized rewards to guide the policy in effectively responding to safety-critical events and understanding real-world causal relationships. To better align with human driving behavior, we incorporate IL into RL training as a regularization term. We introduce a closed-loop evaluation benchmark consisting of diverse, previously unseen 3DGS environments. Compared to IL-based methods, RAD achieves stronger performance in most closed-loop metrics, particularly exhibiting a 3x lower collision rate. Abundant closed-loop results are presented in the supplementary material. Code is available at https://github.com/hustvl/RAD for facilitating future research.

cs.CV↗

arXiv2025

ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving

Recent studies have explored leveraging the world knowledge and cognitive capabilities of Vision-Language Models (VLMs) to address the long-tail problem in end-to-end autonomous driving. However, existing methods typically formulate trajectory planning as a language modeling task, where physical actions are output in the language space, potentially leading to issues such as format-violating outputs, infeasible actions, and slow inference speeds. In this paper, we propose ReCogDrive, a novel Reinforced Cognitive framework for end-to-end autonomous Driving, unifying driving understanding and planning by integrating an autoregressive model with a diffusion planner. First, to instill human driving cognition into the VLM, we introduce a hierarchical data pipeline that mimics the sequential cognitive process of human drivers through three stages: generation, refinement, and quality control. Building on this cognitive foundation, we then address the language-action mismatch by injecting the VLM's learned driving priors into a diffusion planner to efficiently generate continuous and stable trajectories. Furthermore, to enhance driving safety and reduce collisions, we introduce a Diffusion Group Relative Policy Optimization (DiffGRPO) stage, reinforcing the planner for enhanced safety and comfort. Extensive experiments on the NAVSIM and Bench2Drive benchmarks demonstrate that ReCogDrive achieves state-of-the-art performance. Additionally, qualitative results across diverse driving scenarios and DriveBench highlight the model's scene comprehension. All code, model weights, and datasets will be made publicly available to facilitate subsequent research.

cs.CV↗

arXiv2025

UniLat3D: Geometry-Appearance Unified Latents for Single-Stage 3D Generation

High-fidelity 3D asset generation is crucial for various industries. While recent 3D pretrained models show strong capability in producing realistic content, most are built upon diffusion models and follow a two-stage pipeline that first generates geometry and then synthesizes appearance. Such a decoupled design tends to produce geometry-texture misalignment and non-negligible cost. In this paper, we propose UniLat3D, a unified framework that encodes geometry and appearance in a single latent space, enabling direct single-stage generation. Our key contribution is a geometry-appearance Unified VAE, which compresses high-resolution sparse features into a compact latent representation -- UniLat. UniLat integrates structural and visual information into a dense low-resolution latent, which can be efficiently decoded into diverse 3D formats, e.g., 3D Gaussians and meshes. Based on this unified representation, we train a single flow-matching model to map Gaussian noise directly into UniLat, eliminating redundant stages. Trained solely on public datasets, UniLat3D produces high-quality 3D assets in seconds from a single image, achieving superior appearance fidelity and geometric quality. More demos \& code are available at https://unilat3d.github.io/

cs.CV↗

arXiv2025

Few-step Flow for 3D Generation via Marginal-Data Transport Distillation

Flow-based 3D generation models typically require dozens of sampling steps during inference. Though few-step distillation methods, particularly Consistency Models (CMs), have achieved substantial advancements in accelerating 2D diffusion models, they remain under-explored for more complex 3D generation tasks. In this study, we propose a novel framework, MDT-dist, for few-step 3D flow distillation. Our approach is built upon a primary objective: distilling the pretrained model to learn the Marginal-Data Transport. Directly learning this objective needs to integrate the velocity fields, while this integral is intractable to be implemented. Therefore, we propose two optimizable objectives, Velocity Matching (VM) and Velocity Distillation (VD), to equivalently convert the optimization target from the transport level to the velocity and the distribution level respectively. Velocity Matching (VM) learns to stably match the velocity fields between the student and the teacher, but inevitably provides biased gradient estimates. Velocity Distillation (VD) further enhances the optimization process by leveraging the learned velocity fields to perform probability density distillation. When evaluated on the pioneer 3D generation framework TRELLIS, our method reduces sampling steps of each flow transformer from 25 to 1 or 2, achieving 0.68s (1 step x 2) and 0.94s (2 steps x 2) latency with 9.0x and 6.5x speedup on A800, while preserving high visual and geometric fidelity. Extensive experiments demonstrate that our method significantly outperforms existing CM distillation methods, and enables TRELLIS to achieve superior performance in few-step 3D generation.

cs.CV↗

arXiv2025

2D Gaussians Meet Visual Tokenizer

The image tokenizer is a critical component in AR image generation, as it determines how rich and structured visual content is encoded into compact representations. Existing quantization-based tokenizers such as VQ-GAN primarily focus on appearance features like texture and color, often neglecting geometric structures due to their patch-based design. In this work, we explored how to incorporate more visual information into the tokenizer and proposed a new framework named Visual Gaussian Quantization (VGQ), a novel tokenizer paradigm that explicitly enhances structural modeling by integrating 2D Gaussians into traditional visual codebook quantization frameworks. Our approach addresses the inherent limitations of naive quantization methods such as VQ-GAN, which struggle to model structured visual information due to their patch-based design and emphasis on texture and color. In contrast, VGQ encodes image latents as 2D Gaussian distributions, effectively capturing geometric and spatial structures by directly modeling structure-related parameters such as position, rotation and scale. We further demonstrate that increasing the density of 2D Gaussians within the tokens leads to significant gains in reconstruction fidelity, providing a flexible trade-off between token efficiency and visual richness. On the ImageNet 256x256 benchmark, VGQ achieves strong reconstruction quality with an rFID score of 1.00. Furthermore, by increasing the density of 2D Gaussians within the tokens, VGQ gains a significant boost in reconstruction capability and achieves a state-of-the-art reconstruction rFID score of 0.556 and a PSNR of 24.93, substantially outperforming existing methods. Codes will be released soon.

cs.CV↗

arXiv2025

TransLight: Image-Guided Customized Lighting Control with Generative Decoupling

Most existing illumination-editing approaches fail to simultaneously provide customized control of light effects and preserve content integrity. This makes them less effective for practical lighting stylization requirements, especially in the challenging task of transferring complex light effects from a reference image to a user-specified target image. To address this problem, we propose TransLight, a novel framework that enables high-fidelity and high-freedom transfer of light effects. Extracting the light effect from the reference image is the most critical and challenging step in our method. The difficulty lies in the complex geometric structure features embedded in light effects that are highly coupled with content in real-world scenarios. To achieve this, we first present Generative Decoupling, where two fine-tuned diffusion models are used to accurately separate image content and light effects, generating a newly curated, million-scale dataset of image-content-light triplets. Then, we employ IC-Light as the generative model and train our model with our triplets, injecting the reference lighting image as an additional conditioning signal. The resulting TransLight model enables customized and natural transfer of diverse light effects. Notably, by thoroughly disentangling light effects from reference images, our generative decoupling strategy endows TransLight with highly flexible illumination control. Experimental results establish TransLight as the first method to successfully transfer light effects across disparate images, delivering more customized illumination control than existing techniques and charting new directions for research in illumination harmonization and editing.

cs.CV↗

arXiv2025

Snap-Snap: Taking Two Images to Reconstruct 3D Human Gaussians in Milliseconds

Reconstructing 3D human bodies from sparse views has been an appealing topic, which is crucial to broader the related applications. In this paper, we propose a quite challenging but valuable task to reconstruct the human body from only two images, i.e., the front and back view, which can largely lower the barrier for users to create their own 3D digital humans. The main challenges lie in the difficulty of building 3D consistency and recovering missing information from the highly sparse input. We redesign a geometry reconstruction model based on foundation reconstruction models to predict consistent point clouds even input images have scarce overlaps with extensive human data training. Furthermore, an enhancement algorithm is applied to supplement the missing color information, and then the complete human point clouds with colors can be obtained, which are directly transformed into 3D Gaussians for better rendering quality. Experiments show that our method can reconstruct the entire human in 190 ms on a single NVIDIA RTX 4090, with two images at a resolution of 1024x1024, demonstrating state-of-the-art performance on the THuman2.0 and cross-domain datasets. Additionally, our method can complete human reconstruction even with images captured by low-cost mobile devices, reducing the requirements for data collection. Demos and code are available at https://hustvl.github.io/Snap-Snap/.

cs.GR↗

arXiv2025

Turbo-VAED: Fast and Stable Transfer of Video-VAEs to Mobile Devices

There is a growing demand for deploying large generative AI models on mobile devices. For recent popular video generative models, however, the Variational AutoEncoder (VAE) represents one of the major computational bottlenecks. Both large parameter sizes and mismatched kernels cause out-of-memory errors or extremely slow inference on mobile devices. To address this, we propose a low-cost solution that efficiently transfers widely used video VAEs to mobile devices. (1) We analyze redundancy in existing VAE architectures and get empirical design insights. By integrating 3D depthwise separable convolutions into our model, we significantly reduce the number of parameters. (2) We observe that the upsampling techniques in mainstream video VAEs are poorly suited to mobile hardware and form the main bottleneck. In response, we propose a decoupled 3D pixel shuffle scheme that slashes end-to-end delay. Building upon these, we develop a universal mobile-oriented VAE decoder, Turbo-VAED. (3) We propose an efficient VAE decoder training method. Since only the decoder is used during deployment, we distill it to Turbo-VAED instead of retraining the full VAE, enabling fast mobile adaptation with minimal performance loss. To our knowledge, our method enables real-time 720p video VAE decoding on mobile devices for the first time. This approach is widely applicable to most video VAEs. When integrated into four representative models, with training cost as low as $95, it accelerates original VAEs by up to 84.5x at 720p resolution on GPUs, uses as low as 17.5% of original parameter count, and retains 96.9% of the original reconstruction quality. Compared to mobile-optimized VAEs, Turbo-VAED achieves a 2.9x speedup in FPS and better reconstruction quality on the iPhone 16 Pro. The code and models will soon be available at https://github.com/hustvl/Turbo-VAED.

cs.CV↗

arXiv2025

Dynamic 2D Gaussians: Geometrically Accurate Radiance Fields for Dynamic Objects

Reconstructing objects and extracting high-quality surfaces play a vital role in the real world. Current 4D representations show the ability to render high-quality novel views for dynamic objects, but cannot reconstruct high-quality meshes due to their implicit or geometrically inaccurate representations. In this paper, we propose a novel representation that can reconstruct accurate meshes from sparse image input, named Dynamic 2D Gaussians (D-2DGS). We adopt 2D Gaussians for basic geometry representation and use sparse-controlled points to capture the 2D Gaussian's deformation. By extracting the object mask from the rendered high-quality image and masking the rendered depth map, we remove floaters that are prone to occur during reconstruction and can extract high-quality dynamic mesh sequences of dynamic objects. Experiments demonstrate that our D-2DGS is outstanding in reconstructing detailed and smooth high-quality meshes from sparse inputs. The code is available at https://github.com/hustvl/Dynamic-2DGS.

cs.CV↗

arXiv2025

GroundingSuite: Measuring Complex Multi-Granular Pixel Grounding

Pixel grounding, encompassing tasks such as Referring Expression Segmentation (RES), has garnered considerable attention due to its immense potential for bridging the gap between vision and language modalities. However, advancements in this domain are currently constrained by limitations inherent in existing datasets, including limited object categories, insufficient textual diversity, and a scarcity of high-quality annotations. To mitigate these limitations, we introduce GroundingSuite, which comprises: (1) an automated data annotation framework leveraging multiple Vision-Language Model (VLM) agents; (2) a large-scale training dataset encompassing 9.56 million diverse referring expressions and their corresponding segmentations; and (3) a meticulously curated evaluation benchmark consisting of 3,800 images. The GroundingSuite training dataset facilitates substantial performance improvements, enabling models trained on it to achieve state-of-the-art results. Specifically, a cIoU of 68.9 on gRefCOCO and a gIoU of 55.3 on RefCOCOm. Moreover, the GroundingSuite annotation framework demonstrates superior efficiency compared to the current leading data annotation method, i.e., $4.5 \times$ faster than GLaMM.

cs.CV↗

arXiv2025

Genesis: Multimodal Driving Scene Generation with Spatio-Temporal and Cross-Modal Consistency

We present Genesis, a unified framework for joint generation of multi-view driving videos and LiDAR sequences with spatio-temporal and cross-modal consistency. Genesis employs a two-stage architecture that integrates a DiT-based video diffusion model with 3D-VAE encoding, and a BEV-aware LiDAR generator with NeRF-based rendering and adaptive sampling. Both modalities are directly coupled through a shared latent space, enabling coherent evolution across visual and geometric domains. To guide the generation with structured semantics, we introduce DataCrafter, a captioning module built on vision-language models that provides scene-level and instance-level supervision. Extensive experiments on the nuScenes benchmark demonstrate that Genesis achieves state-of-the-art performance across video and LiDAR metrics (FVD 16.95, FID 4.24, Chamfer 0.611), and benefits downstream tasks including segmentation and 3D detection, validating the semantic fidelity and practical utility of the generated data.

cs.CV↗

arXiv2025

From Bench to Bedside: A DeepSeek-Powered AI System for Automated Chest Radiograph Interpretation in Clinical Practice

A global shortage of radiologists has been exacerbated by the significant volume of chest X-ray workloads, particularly in primary care. Although multimodal large language models show promise, existing evaluations predominantly rely on automated metrics or retrospective analyses, lacking rigorous prospective clinical validation. Janus-Pro-CXR (1B), a chest X-ray interpretation system based on DeepSeek Janus-Pro model, was developed and rigorously validated through a multicenter prospective trial (NCT06874647). Our system outperforms state-of-the-art X-ray report generation models in automated report generation, surpassing even larger-scale models including ChatGPT 4o (200B parameters), while demonstrating robust detection of eight clinically critical radiographic findings (area under the curve, AUC > 0.8). Retrospective evaluation confirms significantly higher report accuracy than Janus-Pro and ChatGPT 4o. In prospective clinical deployment, AI assistance significantly improved report quality scores (4.37 vs. 4.11, P < 0.001), reduced interpretation time by 18.5% (P < 0.001), and was preferred by a majority of experts (3 out of 5) in 52.7% of cases. Through lightweight architecture and domain-specific optimization, Janus-Pro-CXR improves diagnostic reliability and workflow efficiency, particularly in resource-constrained settings. The model architecture and implementation framework will be open-sourced to facilitate the clinical translation of AI-assisted radiology solutions.

cs.HC↗

arXiv2025

M3Bench: Benchmarking Whole-body Motion Generation for Mobile Manipulation in 3D Scenes

We propose M3Bench, a new benchmark for whole-body motion generation in mobile manipulation tasks. Given a 3D scene context, M3Bench requires an embodied agent to reason about its configuration, environmental constraints, and task objectives to generate coordinated whole-body motion trajectories for object rearrangement. M3Bench features 30,000 object rearrangement tasks across 119 diverse scenes, providing expert demonstrations generated by our newly developed M3BenchMaker, an automatic data generation tool that produces whole-body motion trajectories from high-level task instructions using only basic scene and robot information. Our benchmark includes various task splits to evaluate generalization across different dimensions and leverages realistic physics simulation for trajectory assessment. Extensive evaluation analysis reveals that state-of-the-art models struggle with coordinating base-arm motion while adhering to environmental and task-specific constraints, underscoring the need for new models to bridge this gap. By releasing M3Bench and M3BenchMaker we aim to advance robotics research toward more adaptive and capable mobile manipulation in diverse, real-world environments.

cs.RO↗

arXiv2025

PixelHacker: Image Inpainting with Structural and Semantic Consistency

Image inpainting is a fundamental research area between image editing and image generation. Recent state-of-the-art (SOTA) methods have explored novel attention mechanisms, lightweight architectures, and context-aware modeling, demonstrating impressive performance. However, they often struggle with complex structure (e.g., texture, shape, spatial relations) and semantics (e.g., color consistency, object restoration, and logical correctness), leading to artifacts and inappropriate generation. To address this challenge, we design a simple yet effective inpainting paradigm called latent categories guidance, and further propose a diffusion-based model named PixelHacker. Specifically, we first construct a large dataset containing 14 million image-mask pairs by annotating foreground and background (potential 116 and 21 categories, respectively). Then, we encode potential foreground and background representations separately through two fixed-size embeddings, and intermittently inject these features into the denoising process via linear attention. Finally, by pre-training on our dataset and fine-tuning on open-source benchmarks, we obtain PixelHacker. Extensive experiments show that PixelHacker comprehensively outperforms the SOTA on a wide range of datasets (Places2, CelebA-HQ, and FFHQ) and exhibits remarkable consistency in both structure and semantics. Project page at https://hustvl.github.io/PixelHacker.

cs.CV↗

arXiv2025

STP4D: Spatio-Temporal-Prompt Consistent Modeling for Text-to-4D Gaussian Splatting

Text-to-4D generation is rapidly developing and widely applied in various scenarios. However, existing methods often fail to incorporate adequate spatio-temporal modeling and prompt alignment within a unified framework, resulting in temporal inconsistencies, geometric distortions, or low-quality 4D content that deviates from the provided texts. Therefore, we propose STP4D, a novel approach that aims to integrate comprehensive spatio-temporal-prompt consistency modeling for high-quality text-to-4D generation. Specifically, STP4D employs three carefully designed modules: Time-varying Prompt Embedding, Geometric Information Enhancement, and Temporal Extension Deformation, which collaborate to accomplish this goal. Furthermore, STP4D is among the first methods to exploit the Diffusion model to generate 4D Gaussians, combining the fine-grained modeling capabilities and the real-time rendering process of 4DGS with the rapid inference speed of the Diffusion model. Extensive experiments demonstrate that STP4D excels in generating high-fidelity 4D content with exceptional efficiency (approximately 4.6s per asset), surpassing existing methods in both quality and speed.

cs.CV↗

arXiv2025

DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving

Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates 10$\times$ reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with the aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions. Code and model will be available at https://github.com/hustvl/DiffusionDrive.

cs.CV↗

arXiv2025

GaussTR: Foundation Model-Aligned Gaussian Transformer for Self-Supervised 3D Spatial Understanding

3D Semantic Occupancy Prediction is fundamental for spatial understanding, yet existing approaches face challenges in scalability and generalization due to their reliance on extensive labeled data and computationally intensive voxel-wise representations. In this paper, we introduce GaussTR, a novel Gaussian-based Transformer framework that unifies sparse 3D modeling with foundation model alignment through Gaussian representations to advance 3D spatial understanding. GaussTR predicts sparse sets of Gaussians in a feed-forward manner to represent 3D scenes. By splatting the Gaussians into 2D views and aligning the rendered features with foundation models, GaussTR facilitates self-supervised 3D representation learning and enables open-vocabulary semantic occupancy prediction without requiring explicit annotations. Empirical experiments on the Occ3D-nuScenes dataset demonstrate GaussTR's state-of-the-art zero-shot performance of 12.27 mIoU, along with a 40% reduction in training time. These results highlight the efficacy of GaussTR for scalable and holistic 3D spatial understanding, with promising implications in autonomous driving and embodied agents. The code is available at https://github.com/hustvl/GaussTR.

cs.CV↗

arXiv2025

Multimodal Mamba: Decoder-only Multimodal State Space Model via Quadratic to Linear Distillation

Recent Multimodal Large Language Models (MLLMs) have achieved remarkable performance but face deployment challenges due to their quadratic computational complexity, growing Key-Value cache requirements, and reliance on separate vision encoders. We propose mmMamba, a framework for developing linear-complexity native multimodal state space models through progressive distillation from existing MLLMs using moderate academic computational resources. Our approach enables the direct conversion of trained decoder-only MLLMs to linear-complexity architectures without requiring pre-trained RNN-based LLM or vision encoders. We propose an seeding strategy to carve Mamba from trained Transformer and a three-stage distillation recipe, which can effectively transfer the knowledge from Transformer to Mamba while preserving multimodal capabilities. Our method also supports flexible hybrid architectures that combine Transformer and Mamba layers for customizable efficiency-performance trade-offs. Distilled from the Transformer-based decoder-only HoVLE, mmMamba-linear achieves competitive performance against existing linear and quadratic-complexity VLMs, while mmMamba-hybrid further improves performance significantly, approaching HoVLE's capabilities. At 103K tokens, mmMamba-linear demonstrates 20.6$\times$ speedup and 75.8% GPU memory reduction compared to HoVLE, while mmMamba-hybrid achieves 13.5$\times$ speedup and 60.2% memory savings. Code and models are released at https://github.com/hustvl/mmMamba

cs.CV↗