AB
AiBoss
project

LATTE3D - A model from NVIDIA for quickly generating 3D objects from text.

LATTE3D is a text-to-3D object generation model developed by NVIDIA researchers. It can quickly generate high-quality 3D content from text prompts (in just 400 milliseconds). The core of this technology lies in a technique called "...

What is LATTE3D?

LATTE3D is a text-to-3D object generation model developed by researchers at NVIDIA's Toronto AI Lab. It can quickly generate high-quality 3D content from text prompts (in just 400 milliseconds). The core of this technology lies in a method called "amortized optimization," which simultaneously optimizes a shared text conditional model across a large number of text prompts. This improves the model's generalization ability to new prompts, thereby reducing the time required to generate each 3D object.

LATTE3D official website entrance

Main functions of LATTE3D

  • Text-to-3D compositing: It can generate corresponding 3D models based on text descriptions. Users can generate 3D objects with specific features and styles by inputting text prompts, such as "a plush amoeba crab wearing a top hat".
  • Quick generation: LATTE3D can generate 3D objects in about 400 milliseconds, which allows it to respond to user input in real time and provide users with instant visual feedback.
  • High-quality rendering: By combining neural fields and textured surface generation, LATTE3D is able to produce textured meshes with a high level of detail, providing visually convincing 3D rendering results.
  • 3D stylization: LATTE3D can also be used as a 3D stylization tool, allowing users to apply new styles or themes to existing 3D assets, thereby creating diverse visual expressions.

LATTE3D's architectural approach

Training LATTE3D consists of two phases: First, volumetric rendering is used to train textures and geometry. To enhance the robustness of the cues, the training objective includes SDS gradients from 3D-aware image priors and a regularized loss that compares a mask of the predicted shape with 3D assets in the library. Then, surface-based rendering is used and textures are trained only to improve quality. Both phases use amortized optimization on a set of cues to maintain fast generation.

LATTE3D's method uses two networks: a texture network T and a geometry network G, both composed of a combination of triplanes and a U-Net. In the first stage, the encoders of both networks share the same weight set. In the second stage, the geometry network G is frozen and the texture network T is updated, and the triplanes are further upsampled using an MLP that embeds the input text.