Pixal3D - A single-image 3D generation project jointly developed by Tencent, Tsinghua University, and other institutions.

What is Pixal3D?

Pixal3D is a single-image 3D generation project launched by Tencent ARC Labs in collaboration with Tsinghua University and Victoria University of Wellington. Pixal3D explicitly upscales pixel features into three-dimensional space through back projection, establishing a direct pixel-to-3D correspondence. It can generate 3D assets with near-reconstruction-level fidelity, featuring fine geometric structures and PBR textures, and is suitable for scenarios such as games, film and television, and digital content creation.

Pixal3D's main functions

• Convert a single image to a high-fidelity 3D model.A single RGB image can generate a 3D asset with detailed geometry and PBR textures.
• Pixel-level 3D correspondence reconstruction: By back-projecting, an explicit mapping between the original pixels and three-dimensional space is established, preserving the local details of the input image.
• PBR material generationSynchronously outputs texture maps with base color, roughness, metallicity, etc., conforming to the physical rendering pipeline.
• Reconstruction-level fidelity outputThe generated results are close to the level of multi-view reconstruction in terms of geometric accuracy and texture consistency.

Pixal3D's technical principles

• Back projection feature enhancementPixal3D's core innovation is to explicitly upscale the pixel features of 2D images to 3D space through a backprojection mechanism. Unlike traditional methods that rely on implicit neural representations or pure optimization, it establishes a direct 3D correspondence for each pixel, accurately mapping the color, edge, and semantic information in the image to the geometric surface.
• Explicit pixel-3D correspondenceThe project constructs an explicit mapping framework that allows 2D pixel coordinates to be directly associated with 3D vertices or surface points. Explicit mapping avoids information loss during intermediate feature distillation, ensuring that the contours and texture details of the generated model are highly consistent with the input image.
• PBR texture and fine geometry co-modelingWhile reconstructing the geometry, Pixal3D employs a joint optimization strategy to generate physically based (PBR) materials. By binding material properties (such as albedo and roughness) to the geometric surface and performing texture mapping using pixel-level correspondences, it achieves high-quality asset output that can be imported into mainstream engines.

How to use Pixal3D

• Visit the project homepage to experience the interactive demo.Visit the Pixal3D Demo at https://huggingface.co/spaces/TencentARC/Pixal3D to access the online experience interface. Upload a single image to load and interactively preview the generated 3D model.
• Download pre-trained model: Obtain the officially released pre-trained weight files.
• Clone code repository configuration environmentGo to the GitHub repository, clone the project to your local machine, and install the dependencies according to the README to ensure that the deep learning framework required for sparse SDF feature volume operations and pixel back projection is supported.
• Prepare input images and perform inferenceSelect an RGB image with a clear foreground and a well-defined subject as input and run the inference script; Pixal3D's image backprojection conditioner will explicitly upscale multi-scale 2D features to 3D feature volume, generating pixel-aligned coarse-structured and high-fidelity detail latent representations.
• Decode and output 3D assetsThe predicted sparse latent representation is decoded into a high-fidelity mesh model via VAE, simultaneously obtaining fine geometry and PBR texture maps, and can be directly exported to industry standard formats.
• Import into the authoring engineImport the generated textured 3D models into mainstream DCC tools and game engines such as Blender, Unity, or Unreal Engine.

Pixal3D's core advantages

• Reconstruction-level fidelityThe generated quality is close to that of multi-view reconstruction, and the geometric and texture details are significantly better than the single-image generation baseline.
• Explicit pixel correspondenceThe back projection mechanism establishes a direct pixel-to-3D mapping, reducing information loss caused by feature compression.
• PBR material natively supportsIt outputs physically rendered textures that meet industry standards in one go, eliminating the need for manual material mapping in post-processing.
• Top academic conferences recognizedThe paper was selected for SIGGRAPH 2026, and the technical solution has undergone rigorous peer review and verification.

Pixal3D project address

• Project official websitehttps://ldyang694.github.io/projects/pixal3d/
• GitHub repositoryhttps://github.com/TencentARC/Pixal3D
• HuggingFace model libraryhttps://huggingface.co/TencentARC/Pixal3D
• arXiv technical paper: https://arxiv.org/pdf/2605.10922

Pixal3D's Competitive Comparison

Applications of Pixal3D

• Rapid production of game assetsGenerate materialized 3D models directly from concept art or reference photos that can be imported into Unity/Unreal Engine.
• Film and Digital Content ProductionIt provides high-fidelity 3D props and basic character models for rehearsals and scene setup.
• E-commerce 3D DisplayAutomatically convert product images into interactive, rotatable 3D display models to enhance the online shopping experience.
• Preview of Architecture and Industrial DesignGenerate 3D prototypes from a single product photo to accelerate the design review and iteration process.