In this paper, we address the challenging task of multimodal reasoning by incorporating the notion of "slow thinking" into multimodal large language models (MLLMs). Our core idea is that models can learn to adaptively use different levels of reasoning to tackle questions of varying complexity. We propose a novel paradigm of Self-structured Chain of Thought (SCoT), which consists of minimal semantic atomic steps. Unlike existing methods that rely on structured templates or free-form paradigms, our method not only generates flexible CoT structures for various complex tasks but also mitigates the phenomenon of overthinking for easier tasks. To introduce structured reasoning into visual cognition, we design a novel AtomThink framework with four key modules: (i) a data engine to generate high-quality multimodal reasoning paths; (ii) a supervised fine-tuning (SFT) process with serialized inference data; (iii) a policy-guided multi-turn inference method; and (iv) an atomic capability metric to evaluate the single-step utilization rate. Extensive experiments demonstrate that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving more than 10% average accuracy gains on MathVista and MathVerse. Compared to state-of-the-art structured CoT approaches, our method not only achieves higher accuracy but also improves data utilization by 5 × and boosts inference efficiency by 85.3%.

We present SeePhys, a large-scale multimodal benchmark for LLM reasoning grounded in physics questions ranging from middle school to PhD qualifying exams. The benchmark covers 7 fundamental domains spanning the physics discipline, incorporating 21 categories of highly heterogeneous diagrams. In contrast to prior works where visual elements mainly serve auxiliary purposes, our benchmark features a substantial proportion of vision-essential problems (75%) that mandate visual information extraction for correct solutions. Through extensive evaluation, we observe that even the most advanced visual reasoning models (e.g., Gemini-2.5-pro and o4-mini) achieve sub-60% accuracy on our benchmark. These results reveal fundamental challenges in current large language models' visual understanding capabilities, particularly in: (i) establishing rigorous coupling between diagram interpretation and physics reasoning, and (ii) overcoming their persistent reliance on textual cues as cognitive shortcuts.Project Page: github.com/SeePhys/seephys-projectHugging Face: huggingface.co/datasets/SeePhys/SeePhys

Recent advances in LVLMs have improved vision-language understanding, but they still struggle with spatial perception, limiting their ability to reason about complex 3D scenes. Unlike previous approaches that incorporate 3D representations into models to improve spatial understanding, we aim to unlock the potential of VLMs by leveraging spatially relevant image data. To this end, we introduce a novel 2D spatial data generation and annotation pipeline built upon scene data with 3D ground-truth. This pipeline enables the creation of a diverse set of spatial tasks, ranging from basic perception tasks to more complex reasoning tasks. Leveraging this pipeline, we construct SPAR-7M, a large-scale dataset generated from thousands of scenes across multiple public datasets. In addition, we introduce SPAR-Bench, a benchmark designed to offer a more comprehensive evaluation of spatial capabilities compared to existing spatial benchmarks, supporting both single-view and multi-view inputs. Training on both SPAR-7M and large-scale 2D datasets enables our models to achieve state-of-the-art performance on 2D spatial benchmarks. Further fine-tuning on 3D task-specific datasets yields competitive results, underscoring the effectiveness of our dataset in enhancing spatial reasoning.

Leveraging diverse robotic data for pretraining remains a critical challenge. Existing methods typically model the dataset’s action distribution using simple observations as inputs. However, these inputs are often incomplete, resulting in a dispersed conditional action distribution—an issue we refer to as coordinate system chaos and state chaos. This inconsistency significantly hampers pretraining efficiency. To address this, we propose 4D-VLA, a novel approach that effectively integrates 4D information into the input to mitigate these sources of chaos. Our model introduces depth and temporal information into visual features with sequential RGB-D inputs, aligning the coordinate systems of the robot and the scene. This alignment endows the model with strong spatiotemporal reasoning capabilities while minimizing training overhead. Additionally, we introduce Memory bank sampling, a frame sampling strategy designed to extract informative frames from historical images, further improving effectiveness and efficiency. Experimental results demonstrate that our pretraining method and architectural components substantially enhance model performance. In both simulated and real-world experiments, our model achieves a significant increase in success rate over OpenVLA.To further assess spatial perception and generalization to novel views, we introduce MV-Bench, a multi-view simulation benchmark. Our model consistently outperforms existing methods, demonstrating stronger spatial understanding and adaptability.

As XR technology continues to advance rapidly, 3D generation and editing are increasingly crucial. Among these, stylization plays a key role in enhancing the appearance of 3D models. By utilizing stylization, users can achieve consistent artistic effects in 3D editing using a single reference style image, making it a user-friendly editing method. However, recent NeRF-based 3D stylization methods encounter efficiency issues that impact the user experience, and their implicit nature limits their ability to accurately transfer geometric pattern styles. Additionally, the ability for artists to apply flexible control over stylized scenes is considered highly desirable to foster an environment conducive to creative exploration. To address the above issues, we introduce StylizedGS, an efficient 3D neural style transfer framework with adaptable control over perceptual factors based on 3D Gaussian Splatting representation. We propose a filter-based refinement to eliminate floaters that affect the stylization effects in the scene reconstruction process. The nearest neighbor-based style loss is introduced to achieve stylization by fine-tuning the geometry and color parameters of 3DGS, while a depth preservation loss with other regularizations is proposed to prevent the tampering of geometry content. Moreover, facilitated by specially designed losses, StylizedGS enables users to control color, stylized scale, and regions during the stylization to possess customization capabilities. Our method achieves high-quality stylization results characterized by faithful brushstrokes and geometric consistency with flexible controls. Extensive experiments across various scenes and styles demonstrate the effectiveness and efficiency of our method concerning both stylization quality and inference speed.

3D reconstruction from multi-view images is a long-standing problem in computer graphic. Neural 3D reconstruction, especially NeuS and its variants, has improved reconstruction quality compared to traditional methods. However, it is still a challenge for these methods to reconstruct fine-grained geometric details since the spherical harmonic positional encoding lacks the ability to express high-frequency signals. In this paper, we propose a multi-resolution tri-plane feature encoding that leverages the detail reconstruction capabilities of high-resolution tri-plane while using the smoothness of low-resolution tri-plane to suppress high-frequency artifacts. Additionally, a progressive training strategy is introduced, gradually merging scene details from coarse to fine granularity, enhancing reconstruction quality while maintaining training stability and reducing difficulty. Furthermore, to address reconstruction challenges arising from sparse viewpoints and inconsistent lighting in image datasets, we introduce normal priors as supervision and propose consistency verification for multi-view normal priors, which assesses the accuracy of normal priors and effectively supervise the reconstructed surfaces. Moreover, we propose a perturbing and fine-tuning strategy on regions of unreliable normal priors to further improve the quality of geometric surface reconstruction.

There has been a high demand for facial makeup transfer tools in fashion e-commerce and virtual avatar generation. Most of the existing makeup transfer methods are based on the generative adversarial networks. Despite their success in makeup transfer for a single image, they struggle to maintain the consistency of makeup under different poses and expressions of the same person. In this article, we propose a robust makeup transfer method which consistently transfers the makeup style of a reference image to facial images in any poses and expressions. Our method introduces the implicit 3D representation, neural radiance fields (NeRFs), to ensure the geometric and appearance consistency. It has two separate stages, including one basic NeRF module to reconstruct the geometry from the input facial image sequence, and a makeup module to learn how to transfer the reference makeup style consistently. We propose a novel hybrid makeup loss which is specially designed based on the makeup characteristics to supervise the training of the makeup module. The proposed loss significantly improves the visual quality and faithfulness of the makeup transfer effects. To better align the distribution between the transferred makeup and the reference makeup, a patch-based discriminator that works in the pose-independent UV texture space is proposed to provide more accurate control of the synthesized makeup. Extensive experiments and a user study demonstrate the superiority of our network for a variety of different makeup styles.

The emergence of 3D Gaussian splatting (3DGS) has greatly accelerated rendering in novel view synthesis. Unlike neural implicit representations like neural radiance fields (NeRFs) that represent a 3D scene with position and viewpoint-conditioned neural networks, 3D Gaussian splatting utilizes a set of Gaussian ellipsoids to model the scene so that efficient rendering can be accomplished by rasterizing Gaussian ellipsoids into images. Apart from fast rendering, the explicit representation of 3D Gaussian splatting also facilitates downstream tasks like dynamic reconstruction, geometry editing, and physical simulation. Considering the rapid changes and growing number of works in this field, we present a literature review of recent 3D Gaussian splatting methods, which can be roughly classified by functionality into 3D reconstruction, 3D editing, and other downstream applications. Traditional point-based rendering methods and the rendering formulation of 3D Gaussian splatting are also covered to aid understanding of this technique. This survey aims to help beginners to quickly get started in this field and to provide experienced researchers with a comprehensive overview, aiming to stimulate future development of the 3D Gaussian splatting representation.

Neural implicit representations, including Neural Distance Fields and Neural Radiance Fields, have demonstrated significant capabilities for reconstructing surfaces with complicated geometry and topology, and generating novel views of a scene. Nevertheless, it is challenging for users to directly deform or manipulate these implicit representations with large deformations in a real-time fashion. Gaussian Splatting (GS) has recently become a promising method with explicit geometry for representing static scenes and facilitating high-quality and real-time synthesis of novel views. However, it cannot be easily deformed due to the use of discrete Gaussians and the lack of explicit topology. To address this, we develop a novel GS-based method (GaussianMesh) that enables interactive deformation. Our key idea is to design an innovative mesh-based GS representation, which is integrated into Gaussian learning and manipulation. 3D Gaussians are defined over an explicit mesh, and they are bound with each other: the rendering of 3D Gaussians guides the mesh face split for adaptive refinement, and the mesh face split directs the splitting of 3D Gaussians. Moreover, the explicit mesh constraints help regularize the Gaussian distribution, suppressing poor-quality Gaussians ( e.g. , misaligned Gaussians, long-narrow shaped Gaussians), thus enhancing visual quality and reducing artifacts during deformation. Based on this representation, we further introduce a large-scale Gaussian deformation technique to enable deformable GS, which alters the parameters of 3D Gaussians according to the manipulation of the associated mesh. Our method benefits from existing mesh deformation datasets for more realistic data-driven Gaussian deformation. Extensive experiments show that our approach achieves high-quality reconstruction and effective deformation, while maintaining the promising rendering results at a high frame rate (65 FPS on average on a single commodity GPU).

The recently proposed neural radiance fields (NeRF) use a continuous function formulated as a multi-layer perceptron (MLP) to model the appearance and geometry of a 3D scene. This enables realistic synthesis of novel views, even for scenes with view dependent appearance. Many follow-up works have since extended NeRFs in different ways. However, a fundamental restriction of the method remains that it requires a large number of images captured from densely placed viewpoints for high-quality synthesis and the quality of the results quickly degrades when the number of captured views is insufficient. To address this problem, we propose a novel NeRF-based framework capable of high-quality view synthesis using only a sparse set of RGB-D images, which can be easily captured using cameras and LiDAR sensors on current consumer devices. First, a geometric proxy of the scene is reconstructed from the captured RGB-D images. Renderings of the reconstructed scene along with precise camera parameters can then be used to pre-train a network. Finally, the network is fine-tuned with a small number of real captured images. We further introduce a patch discriminator to supervise the network under novel views during fine-tuning, as well as a 3D color prior to improve synthesis quality. We demonstrate that our method can generate arbitrary novel views of a 3D scene from as few as 6 RGB-D images. Extensive experiments show the improvements of our method compared with the existing NeRF-based methods, including approaches that also aim to reduce the number of input images.

Neural Radiance Fields (NeRFs) have shown great potential for tasks like novel view synthesis of static 3D scenes. Since NeRFs are trained on a large number of input images, it is not trivial to change their content afterwards. Previous methods to modify NeRFs provide some control but they do not support direct shape deformation which is common for geometry representations like triangle meshes. In this paper, we present a NeRF geometry editing method that first extracts a triangle mesh representation of the geometry inside a NeRF. This mesh can be modified by any 3D modeling tool (we use ARAP mesh deformation). The mesh deformation is then extended into a volume deformation around the shape which establishes a mapping between ray queries to the deformed NeRF and the corresponding queries to the original NeRF. The basic shape editing mechanism is extended towards more powerful and more meaningful editing handles by generating box abstractions of the NeRF shapes which provide an intuitive interface to the user. By additionally assigning semantic labels, we can even identify and combine parts from different objects. We demonstrate the performance and quality of our method in a number of experiments on synthetic data as well as real captured scenes.

3D scene stylization aims at generating stylized images of the scene from arbitrary novel views following a given set of style examples, while ensuring consistency when rendered from different views. Directly applying methods for image or video stylization to 3D scenes cannot achieve such consistency. Thanks to recently proposed neural radiance fields (NeRF), we are able to represent a 3D scene in a consistent way. Consistent 3D scene stylization can be effectively achieved by stylizing the corresponding NeRF. However, there is a significant domain gap between style examples which are 2D images and NeRF which is an implicit volumetric representation. To address this problem, we propose a novel mutual learning framework for 3D scene stylization that combines a 2D image stylization network and NeRF to fuse the stylization ability of 2D stylization network with the 3D consistency of NeRF. We first pre-train a standard NeRF of the 3D scene to be stylized and replace its color prediction module with a style network to obtain a stylized NeRF. It is followed by distilling the prior knowledge of spatial consistency from NeRF to the 2D stylization network through an introduced consistency loss. We also introduce a mimic loss to supervise the mutual learning of the NeRF style module and fine-tune the 2D stylization decoder. In order to further make our model handle ambiguities of 2D stylization results, we introduce learnable latent codes that obey the probability distributions conditioned on the style. They are attached to training samples as conditional inputs to better learn the style module in our novel stylized NeRF. Experimental results demonstrate that our method is superior to existing approaches in both visual quality and long-range consistency.

Implicit neural rendering, especially Neural Radiance Field (NeRF), has shown great potential in novel view synthesis of a scene. However, current NeRF-based methods cannot enable users to perform user-controlled shape deformation in the scene. While existing works have proposed some approaches to modify the radiance field according to the user's constraints, the modification is limited to color editing or object translation and rotation. In this paper, we propose a method that allows users to perform controllable shape deformation on the implicit representation of the scene, and synthesizes the novel view images of the edited scene without re-training the network. Specifically, we establish a correspondence between the extracted explicit mesh representation and the implicit neural representation of the target scene. Users can first utilize well-developed mesh-based deformation methods to deform the mesh representation of the scene. Our method then utilizes user edits from the mesh representation to bend the camera rays by introducing a tetrahedra mesh as a proxy, obtaining the rendering results of the edited scene. Extensive experiments demonstrate that our framework can achieve ideal editing results not only on synthetic data, but also on real scenes captured by users.

We introduce TM-NET, a novel deep generative model for synthesizing textured meshes in a part-aware manner. Once trained, the network can generate novel textured meshes from scratch or predict textures for a given 3D mesh, without image guidance. Plausible and diverse textures can be generated for the same mesh part, while texture compatibility between parts in the same shape is achieved via conditional generation. Specifically, our method produces texture maps for individual shape parts, each as a deformable box, leading to a natural UV map with limited distortion. The network separately embeds part geometry (via a PartVAE) and part texture (via a TextureVAE) into their respective latent spaces, so as to facilitate learning texture probability distributions conditioned on geometry. We introduce a conditional autoregressive model for texture generation, which can be conditioned on both part geometry and textures already generated for other parts to achieve texture compatibility. To produce high-frequency texture details, our TextureVAE operates in a high-dimensional latent space via dictionary-based vector quantization. We also exploit transparencies in the texture as an effective means to model complex shape structures including topological details. Extensive experiments demonstrate the plausibility, quality, and diversity of the textures and geometries generated by our network, while avoiding inconsistency issues that are common to novel view synthesis methods.

We introduce SDM-NET, a deep generative neural network which produces structured deformable meshes. Specifically, the network is trained to generate a spatial arrangement of closed, deformable mesh parts, which respects the global part structure of a shape collection, e.g., chairs, airplanes, etc. Our key observation is that while the overall structure of a 3D shape can be complex, the shape can usually be decomposed into a set of parts, each homeomorphic to a box, and the finer-scale geometry of the part can be recovered by deforming the box. The architecture of SDM-NET is that of a two-level variational autoencoder (VAE). At the part level, a PartVAE learns a deformable model of part geometries. At the structural level, we train a Structured Parts VAE (SP-VAE), which jointly learns the part structure of a shape collection and the part geometries, ensuring the coherence between global shape structure and surface details. Through extensive experiments and comparisons with the state-of-the-art deep generative models of shapes, we demonstrate the superiority of SDM-NET in generating meshes with visual quality, flexible topology, and meaningful structures, benefiting shape interpolation and other subsequent modeling tasks.