Pixal3D is a cutting-edge image-to-3D generation model that transforms 2D images into pixel-perfect 3D assets with unprecedented fidelity. Unlike traditional canonical-space generators that rely on spatial guesswork, Pixal3D explicitly projects pixel features into a 3D volume, ensuring near-reconstruction-level faithfulness to the input view.

Developed by researchers at Tsinghua University and TencentARC, and accepted to SIGGRAPH 2026, Pixal3D represents a significant advancement in the field of 3D generation. The model supports both single and multi-view image input, making it versatile for various use cases including game development, e-commerce product visualization, AR/VR content creation, and 3D printing preparation.

Key features of Pixal3D include its ability to generate complete GLB meshes with PBR (Physically Based Rendering) textures including base color, normal, and roughness maps. This ensures that the generated 3D assets are production-ready and can be directly used in game engines, 3D modeling software, and web-based 3D viewers.

The technology behind Pixal3D is built on PyTorch and leverages the Trellis.2 framework for efficient 3D representation learning. The model is open-source under the MIT license, making it accessible to researchers, developers, and artists worldwide. A Gradio web interface is provided for easy experimentation and usage.

With its ability to generate high-quality 3D assets from simple 2D images, Pixal3D democratizes 3D content creation and enables anyone to create detailed 3D models without specialized skills or expensive software. The project has gained significant traction in the AI and computer vision communities for its innovative approach to 3D generation.