# Luce: relightable 3D assets from one image

Luce (Apple, August 2026) turns a single object image into relightable PBR Gaussians with albedo, metallic-roughness and normals, plus an optional textured mesh. What it outputs, how it compares to SHARP, and where to follow the release.

Canonical URL: https://aiidelist.com/blog/luce

Language: en

Published: 2026-09-24

Updated: 2026-09-24

APPLE LUCE · RELIGHTABLE 3D ASSET GENERATION

Luce is Apple Machine Learning Research’s August 2026 method for image-to-3D asset generation. Instead of a baked-in lighting look, it predicts geometry and physically based materials inside a voxelised multimodal Gaussian cloud, so the result can be relit and exported as a textured mesh.

[Apple Research page](https://machinelearning.apple.com/research/relightable-gaussians-3d-generation)[Compare Luce vs SHARP](/blog/luce-vs-sharp)[Run SHARP locally](/blog/apple-sharp-image-to-3d)

## What “relightable” buys you

Most single-image 3D methods bake the photo’s lighting into colour, so the asset looks wrong under a new light. Luce separates albedo from metallic, roughness and surface normals, which is what game engines and renderers expect from a PBR material. The paper also reports that fine details such as text and logos survive the reconstruction.

## Luce vs SHARP

SHARP turns a photo of a scene into 3D Gaussians for nearby-view synthesis in under a second, and its code is public. Luce targets objects and produces editable, relightable assets. If you want a 3D-photo effect, use SHARP today; if you want a product or game asset, watch Luce.

## Questions

**Can I run Luce today?**

As of the check date we could not confirm a public code or weight release. This page tracks the research; SHARP is the Apple image-to-3D model you can run now.

**What is a Gaussian “cloud” or splat?**

3D Gaussian splatting represents a scene as many small, soft ellipsoids with colour and opacity that render in real time. Luce adds material channels to each Gaussian so lighting can be computed instead of baked.

## At a glance

**Input**

One object image, conditioned on multi-layer features from a pretrained encoder

**Output**

Relightable PBR Gaussians (albedo, metallic-roughness, normals) + optional textured mesh with tangent-space normal maps

**Reported results**

28% lower FID than the strongest baseline on Toys4K; CLIP image alignment 0.8519 vs 0.8299

**Published**

August 2026 · arXiv 2608.23943

**Code & weights**

Not confirmed as released at the time of checking; follow the research page

Checked 2026-09-17

## Official sources

- [Apple Machine Learning Research](https://machinelearning.apple.com/research/relightable-gaussians-3d-generation)
- [Paper (arXiv 2608.23943)](https://arxiv.org/abs/2608.23943)

## Related

- [SHARP demo and setup](/blog/apple-sharp-image-to-3d)
- [Depth Pro](/blog/depth-pro)
- [All Apple models](/blog/apple-ai-models)
- [All Apple models](/blog/apple-ai-models)