When was Pixal3D founded?

Pixal3D was founded in 2026. It is based in Shenzhen, China (Tencent ARC Lab) / Beijing, China (Tsinghua University).

What funding stage is Pixal3D at?

Pixal3D is currently at the Active stage.

Pixal3D

Pixal3D is an academic research project from Tencent ARC Lab, Tsinghua University, and Victoria University of Wellington — an open-source image-to-3D generation model using pixel-aligned back-projection for high-fidelity 3D asset creation from single or multi-view images. Pixal3D was founded in 2026. Based in Shenzhen, China (Tencent ARC Lab) / Beijing, China (Tsinghua University).

Founded: 2026
Headquarters: Shenzhen, China (Tencent ARC Lab) / Beijing, China (Tsinghua University)

Value proposition

Pixal3D generates high-fidelity 3D assets (GLB meshes with PBR textures) from a single image using pixel-aligned back-projection, achieving near-reconstruction-level fidelity by establishing direct pixel-to-3D correspondences instead of loose attention-based conditioning.

Products and solutions

Pixal3D open-source model (SIGGRAPH 2026); Gradio web demo on Hugging Face Spaces; Command-line inference script for GLB mesh export; Two branches: main (TRELLIS.2 backbone, improved) and paper (Direct3D-S2 backbone, reproducing paper results)

Unique value

Pixel-aligned 3D generation that explicitly lifts multi-scale image features into a 3D feature volume via back-projection, creating direct pixel-to-3D correspondence — achieving 93.57 IoU on Toys4K single-view evaluation, significantly outperforming Hunyuan3D-2.1 (83.33), TRELLIS (79.48), and TripoSG (73.54).

Target customer

Academic researchers in 3D computer vision and graphics; Game developers and indie studios; E-commerce and AR content creators; 3D artists needing fast asset prototyping

Industries served

Gaming; E-commerce; Augmented Reality; 3D Content Creation; Computer Vision Research

Technology advantage

Pixel-aligned structured latent representation learning via VAE; Image back-projection-based conditioner for explicit 2D-3D correspondence; Two-stage generative process (coarse structure + detailed latents); Natural extension to multi-view generation; Modular pipeline for object-separated 3D scenes; Built on TRELLIS.2 backbone

How they differentiate

Unlike most 3D-native generators that synthesize shapes in canonical space and inject image cues via attention (leaving pixel-to-3D associations ambiguous), Pixal3D directly generates 3D in a pixel-aligned way consistent with the input view, using pixel back-projection conditioning that explicitly lifts multi-scale image features into a 3D feature volume.

Main competitors

Hunyuan3D-2.1 (Tencent); TRELLIS / TRELLIS.2 (Microsoft); TripoSG (Tripo AI); Direct3D-S2

Key partnerships

Tencent ARC Lab (primary research lab); Tsinghua University (BNRist, Department of Computer Science and Technology); Victoria University of Wellington; Hugging Face (model hosting and demo Spaces)

Major milestones

April 2026: Paper accepted to SIGGRAPH 2026; May 2026: Inference code and online Gradio demo released on Hugging Face; May 2026: Improved main branch based on TRELLIS.2 backbone released; May 2026: arXiv paper 2605.10922 published; May 2026: License changed to MIT (commercial use permitted); May 2026: Training code and data preparation toolkit released; ~1,300 GitHub stars, ~110 forks

Growth metrics

~1,300 GitHub stars; ~110 forks; SIGGRAPH 2026 publication; 93.57 IoU on Toys4K benchmark; 4.91/5 user-study fidelity score; MIT license

Market positioning

Academic research project positioned as a fidelity-focused alternative to existing image-to-3D methods, demonstrating state-of-the-art pixel-level faithfulness (93.57 IoU on Toys4K) through explicit pixel-to-3D correspondence, but restricted to academic/non-commercial use under its license.

Geographic focus

Global (open-source research project); primarily China-based (Tencent ARC Lab + Tsinghua University)

Official website: https://ldyang694.github.io/projects/pixal3d/