Apple Luce Generates Relightable PBR 3D Assets From a Single Image

Apple researchers introduce Luce for production-oriented image-to-3D generation

Apple Machine Learning Research published Luce: Relightable Gaussians for 3D Asset Generation in August 2026, with the paper submitted to arXiv on August 25, 2026.

Luce targets a practical limitation in image-to-3D systems: generating something that looks plausible from several views is not the same as generating an asset that can be relit and used in a conventional graphics pipeline.

The method produces a representation containing both geometry and physically based rendering (PBR) material information, including albedo, metallic-roughness and surface normals.

A multimodal Gaussian representation carries geometry and materials

Luce represents an object as a voxelized multimodal Gaussian cloud. Dedicated Gaussian primitives encode different PBR modalities rather than collapsing appearance into a single view-dependent representation.

A variational autoencoder compresses those modalities into a unified material-aware latent space. A rectified-flow transformer then generates that latent from one input image using multi-layer features from a pretrained image encoder.

The decoder can produce relightable PBR Gaussians and, optionally, a textured mesh with a tangent-space normal map. That optional mesh output is important for workflows that still depend on standard asset formats and downstream rendering tools.

The paper reports stronger image-to-3D benchmark results

On Toys4K, the authors report a 28% FID improvement over the strongest baseline in their comparison.

They also introduce an evaluation set made from AI-generated images. On that benchmark, Luce reaches a 0.8519 CLIP image-alignment score, compared with 0.8299 for the best baseline reported in the paper.

Those are author-reported results on the stated evaluation setups. They do not establish that Luce will outperform every commercial or research image-to-3D system on every asset category.

Why relightability matters

A 3D asset intended for games, simulation, spatial computing, product visualization or robotics often needs to behave consistently under lighting conditions that were not present in the source image.

By explicitly generating PBR material channels along with geometry, Luce is aimed at the gap between visually convincing 3D generation and assets that can participate in standard rendering pipelines.

The paper also emphasizes preservation of fine details such as text, logos and inscriptions, which are common failure points in generative 3D systems.

What is released and what is not established

Released: the Apple research page and public paper describe the Luce method and evaluation results.

Not established by the announcement: a consumer Apple product, a public hosted Luce generation service, production guarantees for arbitrary assets, or universal superiority over all existing 3D-generation systems.

Luce is best understood as a newly published research system for more material-aware, relightable single-image 3D generation.