Pixal3DImage-to-3D generator that lifts pixel features into 3D for view-consistent, PBR-ready GLB assets

Pixal3D is an image-to-3D generation tool that converts single or multi-view 2D images into textured 3D assets exportable in GLB format. It originates from research by groups at Tsinghua University and TencentARC, and the project is positioned around an academic contribution accepted to SIGGRAPH 2026. Code is published on GitHub with models hosted on Hugging Face, and a hosted web demo lets users try it without a local setup. The tool's central claim is that it tackles the 2D-to-3D correspondence problem differently from many 3D-native generators. Rather than predicting a shape in a generic canonical pose using loose attention, Pixal3D uses an explicit "pixel back-projection" scheme that lifts multi-scale 2D image features into a 3D feature volume. The intended result is view-consistent output where the generated front matches the input image closely, preserving proportions, silhouettes, and fine detail instead of producing a plausible-but-invented model. Generation is built on the Trellis.2 backbone, which the project credits for inference speed and feature extraction quality. Outputs are meshes equipped with Physically Based Rendering materials—base color, normal, and roughness maps—so assets can drop into engines like Unity and Unreal, or into Blender, without manual UV work or texture painting. The pipeline scales from a single image up to multiple views: with several angles it aggregates back-projected features to improve 360-degree consistency and fill in occluded regions, which suits character turnaround sheets. Beyond isolated props, Pixal3D advertises a modular pipeline capable of parsing complex images into object-separated 3D scenes, aimed at rapid environment prototyping rather than only single-object generation. The target users are technical and 3D artists, game developers, and spatial-computing creators who need outputs that faithfully match reference or concept art. As an open-source, research-backed project, Pixal3D's main differentiator is fidelity to the input view rather than raw creative variety—an emphasis that competes with tools like Meshy, Tripo, Rodin, and Trellis itself. Prospective users should weigh that its strongest results center on matching a given image, and that, as a relatively new research release, real-world robustness across diverse inputs, occluded geometry, and full scene synthesis is best verified against your own assets. Specific pricing, usage limits, and commercial-licensing terms are not detailed in the available material.