Apple Research Introduces Luce for Relightable Single-Image 3D Generation

Apple researchers publish Luce for relightable 3D generation

Apple Machine Learning Research has published Luce: Relightable Gaussians for 3D Asset Generation, a research system for turning a single image into a 3D asset whose appearance can be relit under new illumination.

The paper was submitted to arXiv on August 25, 2026. This is a research result, not an announced Apple product or generally available model.

Luce targets a limitation of many image-to-3D systems: they can reproduce an object's appearance from the reference view, but the generated representation may not separate material properties well enough for realistic relighting or integration into conventional rendering pipelines.

What Luce changes

The researchers represent an object as a voxelized multimodal Gaussian cloud. Separate Gaussian primitives encode geometry-related and physically based rendering information such as albedo, metallic-roughness and surface normals.

A variational autoencoder compresses those modalities into a shared material-aware latent space. A rectified-flow transformer then predicts that latent representation from a single input image, using multi-layer features from a pretrained image encoder to preserve both semantic information and fine spatial detail.

The decoded result can be rendered as relightable PBR Gaussians and can optionally be converted into a textured mesh with a tangent-space normal map. That makes the work relevant to graphics pipelines where generated assets need to respond to new lighting rather than behave like view-dependent image billboards.

Reported results

On the Toys4K benchmark, the Apple research page reports that Luce improves FID by 28% over the strongest baseline used in the study. The team also evaluates on a benchmark of AI-generated reference images and reports a CLIP image-alignment score of 0.8519, compared with 0.8299 for the strongest baseline in that comparison.

The researchers emphasize preservation of fine details such as text, logos and inscriptions, alongside geometry and material properties.

These are paper-reported results and should not be treated as independent production benchmarks. Performance in a real asset pipeline will also depend on input quality, object category, topology requirements, relighting conditions, runtime and downstream editing needs.

Why it matters

Generative 3D systems are increasingly useful for games, spatial computing, product visualization, simulation and synthetic-data creation. For many of those workflows, producing a visually plausible object is only the first step. Artists and rendering systems also need material information that behaves consistently when lights, cameras or environments change.

Luce is notable because it treats relightability and material structure as first-class outputs of generation rather than a later reconstruction step. The approach also connects recent progress in rectified-flow generative models with Gaussian-based 3D representations.

For developers, the important status distinction is that Apple has published the research and paper, but has not announced a public Luce API, downloadable checkpoint or product integration on the cited research page. The work is therefore best viewed as a research direction for material-aware image-to-3D generation rather than a shipping Apple service.