Pixal 3D

Pixal3D is a pixel-aligned 3D generation paradigm that transforms 2D images into high-fidelity 3D assets. Developed by researchers at Tsinghua University and TencentARC, it explicitly lifts multi-scale 2D image features directly into a 3D feature volume via back-projection, establishing direct pixel-to-3D correspondence. This approach eliminates spatial guesswork and preserves exact proportions, silhouettes, and micro-details from the original concept art.

Key Features

• Direct Pixel-to-3D Correspondence: Locks in every pixel to establish unambiguous correspondence, eliminating AI hallucinations on the primary view.
• Production-Ready PBR Materials: Generates high-resolution meshes equipped with Physically Based Rendering (PBR) textures, including base color, normal, and roughness maps.
• Multi-View Aggregation: Scales from a single image to multi-view character sheets, seamlessly combining back-projected features for superior topological accuracy and 360-degree consistency.
• Modular Scene Synthesis: Parses complex 2D scene structures to generate high-fidelity, object-separated 3D scenes.
• Trellis.2 Backbone: Powered by the highly optimized Trellis.2 architecture for state-of-the-art inference speeds and quality.

Workflow

1. Upload a single image or multi-view shots.
2. The architecture explicitly lifts 2D pixels into a 3D feature volume.
3. The backbone synthesizes high-resolution meshes and applies detailed PBR textures.
4. Export the result immediately as a standard GLB asset.

Who It’s For

Pixal3D is designed for 3D artists, game developers, spatial computing creators, and researchers who demand absolute visual faithfulness and production-ready assets for engines like Unity, Unreal, and Blender.