Pixal3D converts single images to high-fidelity 3D geometry with pixel back-projection
Pixal3D, a locally runnable model from TencentARC, generates high-fidelity 3D assets from single images by explicitly back-projecting pixel features into 3D space rather than using attention-based injection.
Pixal3D, TencentARC's locally runnable 3D generator, converts single images into high-fidelity geometry and PBR textures by back-projecting pixel features directly into 3D space. Unlike prior methods that inject image features through attention mechanisms, Pixal3D establishes explicit pixel-to-3D correspondences, producing near-reconstruction-level detail. The weights are available on HuggingFace.
Early adopters report cleaner surface geometry than Trellis.2, though formal benchmarks remain unpublished. At least one user successfully ran it on an RTX 5090 after resolving dependency issues; the model card does not yet list VRAM requirements or recommended hardware specs. TencentARC has also deployed an online demo for users without local hardware. The open-weight nature means practitioners can fine-tune or integrate it into local pipelines without API restrictions—a key advantage for custom 3D workflows.