Apple has moved LiTo from research paper to usable code, giving developers a clearer path to test 3D AI just as spatial computing needs better content tools.
Apple's most interesting AI move right now is not another chatbot feature. It is LiTo, a public image-to-3D research project that can turn a single image into a realistic 3D object while preserving how light, reflections and materials behave from different angles.
That matters because 3D asset creation is still one of the slowest parts of building games, product visualizations, robotics simulations and spatial apps. Text and image generation have already changed creative workflows. 3D generation is where that pressure moves next, and Apple putting code in public view gives builders something more useful than a paper abstract.
Apple's project page and GitHub repository show LiTo, short for Surface Light Field Tokenization, was accepted to ICLR 2026 and built by Jen-Hao Rick Chang, Xiaoming Zhao, Dorian Chan and Oncel Tuzel. The paper first appeared on arXiv in March 2026, but the developer story is current because the public repository has recently been updated with code, demos, configuration files, data splits and model licensing details. Pinokio's repository index showed the Apple ml-lito project updated on May 19 and indexed on May 23, with recent commits adding code, cleaning the environment and fixing submodules.
The core idea behind LiTo is straightforward, even if the implementation is not. Most 3D reconstruction systems try to recover shape first, then attach surface color or texture later. That can work well for matte objects, but it often struggles with anything shiny or glossy, because the object changes appearance as the viewer moves.
LiTo tries to solve that by encoding geometry and view-dependent appearance together in a shared latent 3D representation. In plain terms, it does not only ask what the object looks like. It asks how the object should keep looking as light catches it from another angle. Apple says the method can reproduce effects such as specular highlights and Fresnel reflections under complex lighting, which are exactly the details that make generated 3D assets feel less like cutouts.
The repository also gives developers practical numbers to work with. Apple lists image-to-3D generation at 4.7 seconds on an Nvidia H100 after torch compile, and says Linux systems with Nvidia A100, H100 and B200 GPUs get full support for training, the interactive image-to-3D demo and tokenizer notebooks. Mac users can run the interactive image-to-3D demo through MLX on Apple Silicon, but the tokenizer notebook still requires Linux with an Nvidia GPU.
That split is important. LiTo is open to developers, but it is not yet a push-button production tool for every laptop. The serious training and experimentation path still belongs to teams with strong GPU access. For startups, that means LiTo is immediately useful for prototyping, benchmarking and pipeline research, while broader commercial use will depend on tooling around it.
The Open 3D Field Is Getting Crowded
Apple is not entering an empty market. Microsoft's TRELLIS project has become a major reference point for native 3D generation, with support for text or image conditions. Tencent's Hunyuan3D 2.1 has pushed open 3D asset generation toward physically based rendering materials, while Roblox's Cube 3D connected generative models directly to creator workflows inside Roblox Studio.
LiTo's pitch is more specific. It is not trying to be the broadest creator product on day one. It is trying to improve fidelity, alignment and lighting consistency from an input image. Apple's comparison page highlights TRELLIS as a baseline and notes that TRELLIS can sometimes output objects in an incorrect orientation because it does not respect the same camera coordinate system. LiTo's repository lists alignment with the image as one of its features.
That may sound like a small technical detail, but it has real workflow consequences. If a game studio, design startup or enterprise AR team has to manually correct orientation, lighting and material behavior every time a generated asset enters a pipeline, the labor savings shrink quickly. Better alignment means generated objects have a higher chance of being useful before a human artist cleans them up.
The data section also shows how Apple is thinking about reproducibility. The repository says its train, validation and test splits cover Objaverse and ObjaverseXL, with 84,825 training samples from Objaverse and 155,275 from ObjaverseXL, for a total of 240.1k training samples after filtering. That gives outside researchers a clearer path to compare results instead of relying only on polished demo clips.
What This Means For Vision Pro Builders
The obvious question is whether LiTo is part of a larger Apple developer strategy around spatial computing. Apple has not framed it that way directly, and a research repository is not the same thing as a Vision Pro product roadmap. Still, the connection is hard to ignore.
Vision Pro and future spatial devices need more 3D content than traditional app stores ever required. Every startup building training simulations, digital twins, commerce previews, education apps or immersive collaboration tools eventually runs into the same bottleneck: 3D assets are expensive, slow and hard to scale. A better image-to-3D model does not remove artists from the process, but it can move the first draft from days to minutes.
Apple's public release also fits a broader pattern. The company has been more willing to publish machine learning tools and research code when it helps developers work closer to Apple hardware, especially through MLX on Apple Silicon. LiTo's Mac demo support is limited, but it gives builders a reason to test 3D AI locally instead of sending every asset idea to a cloud API.
The next test is adoption. If developers wrap LiTo into Blender plugins, asset review tools, game prototyping workflows or Vision Pro pipelines, Apple gets more than citations. It gets a developer community learning its 3D assumptions before spatial computing has truly gone mainstream.
For now, LiTo should be read as a serious research release with practical hooks, not a finished creative suite. But that is enough to matter. The 3D AI race is moving from impressive demos to usable infrastructure, and Apple has just made clear that it wants a place in that stack.
Also read: Solana's hackathon surge shows builders are ignoring SOL's slump • Apple is using fraud numbers to defend its App Store control • Solana builders are moving faster than the SOL price suggests
