The LiTo model combines geometry and lighting to deliver more lifelike 3D reconstructions than traditional methods.
Built on latent space, it compresses and rebuilds visual data using an encoder-decoder system.
Early tests suggest it outperforms models like TRELLIS, with potential use in gaming, VR, and design.
Apple researchers have announced a new AI model that can reconstruct a 3D object from a single image while accurately preserving lighting effects such as reflections and highlights at various viewing angles. Apple’s machine learning website features research titled LiTo: Surface Light Field Tokenisation, which proposes a method that goes beyond traditional 3D reconstruction. While most existing systems concentrate on shape or basic surface appearance, this method combines geometry and how light interacts with the object, making the final result look much more realistic.
SurveyThe model is based on the concept of latent space. Simply put, the system compresses visual data into mathematical representations. This enables it to understand not only the structure of an object, but also how it should appear in various lighting and viewing conditions.
This process actually consists of two steps. First, an encoder converts the image into a compact latent representation. After that, a decoder reconstructs the object in 3D, including all the details like reflections and lighting shifts based on angle. This allows the model to make real-world visual effects, which are usually difficult to capture.
To train the system, researchers used thousands of objects rendered from various angles and lighting configurations. Instead of processing all of the data at once, the model learned from smaller samples and gradually improved its ability to predict full 3D structures and lighting behaviour.
Once it gets trained, the system will take a single image and create a full 3D model of the object with all the dynamic lighting effects. Early comparisons, as per Apple, show that it outperforms existing models such as TRELLIS in producing realistic results. If true, this can be really handy for many tasks, including gaming, virtual reality, and product design.
Follow Us
Ashish Singh
Ashish Singh is the Chief Copy Editor at Digit. He's been wrangling tech jargon since 2020 (Times Internet, Jagran English '22). When not policing commas, he's likely fueling his gadget habit with coffee, strategising his next virtual race, or plotting a road trip to test the latest in-car tech. He speaks fluent Geek. View Full Profile
