What is it
Pixal3D is an AI-powered, pixel-aligned 3D generation platform designed to transform 2D concepts into production-ready 3D assets in minutes. Developed by researchers from Tsinghua University in collaboration with Tencent ARC, Pixal3D addresses the 2D-to-3D correspondence bottleneck by lifting pixel features directly into a 3D feature volume, delivering reconstruction-level fidelity, intricate geometry, and PBR textures that are ready for real-time engines. The system supports single-image and multi-view inputs, producing view-consistent geometry that exports as standard GLB files with Physically Based Rendering (PBR) materials. Pixal3D is open-source-friendly, with code available on GitHub and models hosted on Hugging Face, plus a browser-based Gradio demo for quick experimentation. Its Trellis.2 backbone and modular pipeline enable scalable, production-ready asset generation—from isolated props to complex scenes.
Features
- Pixel-aligned 3D generation with direct pixel-to-3D correspondence
- Reconstruction-level fidelity: every pixel is lifted into 3D, preserving silhouettes and proportions
- Pixel Back-Projection: explicit, back-projected 2D features form a precise 3D feature volume
- Trellis.2 backbone: optimized, fast, high-quality inference for high-resolution meshes and textures
- Multi-view aggregation: seamless fusion of back-projected features from multiple angles for superior 360-degree accuracy
- View-consistent, production-ready output: front views align precisely with reference imagery
- PBR textures and materials: export fully textured GLB assets with Base Color, Normal, and Roughness maps
- GLB asset export: ready-to-use 3D models compatible with Unity, Unreal, Blender and other engines
- Scene and object separation: modular pipeline capable of reconstructing complex, multi-object environments
- Open-source access: GitHub repository, Hugging Face model hosting, Gradio web demo, extensive docs
- Industry-grade validation: research accepted to SIGGRAPH 2026; designed for professional artists, developers, and researchers
- Scalable inputs: effective from a single image to multi-view character sheets or scenes
- Occlusion completion and detail preservation: fills in hidden areas with plausible, faithful geometry
How to Use
- Start with reference images: upload a single image or multiple views of your character, prop, or scene element
- Pixel back-projection: Pixal3D explicitly lifts 2D pixels into a 3D feature volume, establishing accurate spatial relationships
- Generate geometry and textures: the Trellis.2 backbone synthesizes high-resolution meshes and applies detailed PBR textures aligned to your views
- Export and deploy: download high-fidelity GLB assets that are ready for immediate use in your game engine or 3D software
- Accessibility and experimentation: try the browser-based Gradio demo on Hugging Face; for local runs, use the GitHub repository (modern GPU recommended)
- Output formats and pipeline compatibility: GLB with PBR maps (Base Color, Normal, Roughness); plug-and-play with Unity, Unreal, Blender
Pricing
- Free-to-try access: Pixal3D offers a no-cost browser demo and a free trial experience to explore core capabilities
- Open-source and community access: code and models are available on GitHub and Hugging Face for integration into custom pipelines
- Enterprise and extended use: exact pricing for ongoing, production-scale usage is typically available on request or via partnerships; check the official site or contact the team for details
Tips
- Use multiple views when possible: multi-view inputs significantly improve 360-degree quality and topological accuracy
- Ensure reference images are clean and aligned: consistent lighting and clear silhouettes help preserve original proportions
- Leverage production-ready textures: rely on the included PBR maps to reduce manual UV/texturing work
- Validate front-view fidelity: Pixal3D’s pixel-aligned approach minimizes hallucinations; compare output against the reference to confirm fidelity
- Integrate into existing pipelines: export GLB assets with standard materials to minimize tweaks in Unity, Unreal, or Blender
- Experiment with scene complexity: Pixal3D can handle multi-object environments; start with prop-level assets and scale to scenes as needed
- Hardware and access: while a browser demo runs without installation, a modern GPU is recommended for local runs via GitHub
Frequently Asked Questions
- What is Pixal3D? A pixel-aligned 3D generation paradigm that creates high-fidelity 3D assets directly from 2D images, preserving pixel-level faithfulness and exact silhouettes.
- How is it different from other Image-to-3D AIs? Unlike many 3D-native generators that synthesize in a canonical space, Pixal3D lifts multi-scale 2D features into a 3D volume via back-projection, ensuring precise pixel-to-3D correspondence and fewer distortions.
- Can I use multiple reference images? Yes. Pixal3D supports multi-view inputs and aggregates back-projected features from different angles for improved 3D accuracy.
- What formats are exported? Output is GLB, with geometry plus PBR texture maps (Base Color, Normal, Roughness) ready for immediate use.
- Can it generate full scenes, not just individual props? Yes. The modular pipeline can reconstruct complex, object-separated 3D scenes from imagery.
- Who developed Pixal3D? Dong-Yang Li and researchers from Tsinghua University in collaboration with Tencent ARC; the work was accepted to SIGGRAPH 2026.
- Do I need powerful hardware? A modern GPU is recommended for local runs; however, you can try the browser-based Gradio demo without installation.
- What makes it suitable for production workflows? It delivers production-ready GLB assets with PBR materials and view-consistent geometry, reducing the need for manual retopology and texture painting.
- Where can I learn more or try it? Access the Gradio demo on Hugging Face, view the GitHub repository for code, and explore the Hugging Face model hosting for integration details.
This comprehensive overview highlights Pixal3D’s unique value: pixel-aligned, view-consistent 3D generation that merges cutting-edge research with practical, production-ready assets. By converting 2D concepts into high-fidelity GLB models with PBR textures, Pixal3D streamlines workflows for 3D artists, game developers, and spatial computing creators seeking accurate, ready-to-deploy assets without the typical hiccups of traditional 3D reconstruction.