Pixal3D is a state-of-the-art image-based 3D asset generation solution developed by Tencent ARC Lab and released in May 2026. It can simultaneously reconstruct high-resolution geometric meshes and realistic textures from a single 2D image. This system utilizes Microsoft's next-generation O-Voxel-based 4B parameter framework, the TRELLIS.2 Flow-matching Transformer architecture, as its backbone, enabling it to rapidly generate complex physically-based rendering (PBR) materials and intricate topological structures that were previously difficult for existing models to represent. Notably, by fully open-sourcing the entire training code, data pre-processing pipeline, and all weights for local inference and web demos under a complete MIT license, it supports researchers and developers in leveraging the latest 3D vision infrastructure and freely building custom pipelines without license restrictions. Conventional image-to-3D generation methods typically generate 3D shapes in a canonical space and indirectly inject 2D image information through cross-attention, resulting in a significant limitation where the accurate pixel mapping (Pixel-to-3D correspondence) between the 2D input image and the 3D shape is inconsistent. To illustrate, if existing 3D generators are like virtual sculptors who glance at a 2D photograph of an object and rely on their memory to mold clay, then Pixal3D is like a meticulous restorer who fixes the original photograph tightly behind a transparent glass plate and traces the pixels' outlines and textures, leaving marks on the three-dimensional surface. Pixal3D designs a Pixel Back-projection Conditioning technique to directly define the 3D space in a way that completely matches the input view, thereby achieving high-precision pixel alignment, where the generated result perfectly overlaps with the input 2D image, ensuring a level of precision that approaches actual reconstruction. In the fields of biotechnology and bio-image analysis, Pixal3D is used as a digitization tool to instantly convert 2D stereoscopic images and projected images into 3D stereoscopic shapes for research purposes. When a single transmission electron microscopy (TEM) image or a complex biological tissue section image, captured in a sterile experimental animal or micro-analysis device environment, is input into this engine, a 3D mesh file with geometric stereoscopic curvature and surface texture is rapidly and automatically generated. These outputs can be combined with multi-view data to be processed into high-resolution volumetric data at the cellular level, or directly transferred to biological visualization software such as Blender and PyMOL, and can be used as a key source for analyzing the structural interactions of cell membrane organelles, outputting educational stereoscopic printing assets, or designing structural models for research on drug complex binding interfaces.

Pixal3D is a state-of-the-art image-based 3D asset generation solution developed by Tencent ARC Lab and released in May 2026. It can simultaneously reconstruct high-resolution geometric meshes and realistic textures from a single 2D image. This system uses Microsoft's next-generation O-Voxel-based 4B parameter framework, the TRELLIS.2 Flow-matching Transformer architecture, as its backbone, enabling it to represent complex physically-based rendering (PBR) materials and stereoscopic images that were difficult for previous models to handle.

3D Reconstruction of Cellular Organelles from a Single Microscopy Image: Input a single 1024x1024 resolution 2D transmission electron microscopy (TEM) image of mitochondria. Using Pixal3D's inverse projection algorithm and the TRELLIS.2 backbone, generate a 3D mesh on the GPU within 3-5 seconds. The generated OBJ mesh can be combined with Blender and PyMOL to support cell membrane curvature measurement, volume analysis, and the creation of 3D printing educational materials.