
Apple's Luce generates relightable 3D assets from a single image.
It captures geometry and PBR materials, improving FID by 28% over the strongest baseline.
It preserves fine details like text and logos.
What happened
Apple researchers introduced Luce, a 3D representation that combines geometry and PBR materials into a voxelized multimodal Gaussian cloud. It generates 3D assets from a single image, using a rectified-flow transformer to create a latent representation.
Why it matters
Luce achieves state-of-the-art single-image-to-3D generation on the Toys4K benchmark, improving FID by 28% over the strongest baseline. It also outperforms the best baseline on a new benchmark of AI-generated images, with a CLIP image-alignment score of 0.8519 vs. 0.8299.
What to watch
Luce preserves fine details like text, logos, and inscriptions, and it can output either relightable PBR Gaussians or an optional textured mesh with tangent-space normal map. This could make AI-generated 3D assets more usable in standard rendering pipelines.
Ask the AI about this article →
The Luce paper from Apple addresses a key limitation in 3D asset generation: producing models that integrate into standard rendering pipelines. By unifying geometry and PBR materials in a single representation, it enables relighting, a requirement for realistic rendering in games and film. The use of a variational autoencoder and rectified-flow transformer is a technical approach to compress and generate the latent space, but the practical result is more important: assets with preserved fine details like text and logos, which are often problematic for other methods.
The improvement of FID by 28% on Toys4K marks a notable advance in single-image-to-3D generation quality. The introduction of a benchmark for AI-generated images may also provide a new standard for evaluating such models. For businesses, this technology could lower the barrier to creating 3D content, potentially impacting industries like e-commerce, virtual reality, and product design, though the paper does not discuss commercial applications. The optional textured mesh output suggests compatibility with existing workflows, which might ease adoption.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
AI-generated animal videos and pictures, from polar bear rescues to found pets, are so prevalent that viewers…

Netflix's upcoming film 'Hamnet' has been accused of using AI due to an apparent error in a promotional image…

Stability AI, the startup behind Stable Diffusion, raised $76 million in Series B funding, bringing its total…

Apple researchers have introduced STARFlow2, a new model built on the Pretzel architecture that combines a fro…

Asics announced it will begin offering an AI-powered bear-detection app (AI-App クマ検知) starting September 1

A software engineer created By-Its-Cover, a recommendation system that suggests books based on their cover ima…
