Meta releases new 3D asset generation AI system, claiming to achieve "qualitative leap"

Meta recently announced a new version of its 3D asset generation AI system, calling it a "significant technological leap." The system, called AssetGen 2.0, aims to address the details of previous models when generating 3D content, providing key support for creators to create more realistic virtual worlds.

From “usable” to “easy to use”: addressing the pain points of 3D generation

The Horizon desktop editor (a tool for building the Horizon Worlds virtual world) launched last year has implemented the AI-generated 3D asset function based on text prompts - users do not need professional modeling skills and can build a custom virtual space for free in just a few hours. However, early AI-generated assets have significant quality issues, especially when viewed up close in VR devices.

 The core mission of AssetGen 2.0 is to overcome this industry challenge.

Meta claims that the new version achieves "breakthrough improvements in detail and fidelity" and the generated 3D models have "geometric consistency with fine details", setting a new benchmark for the industry and "pushing the boundaries of what is possible with generative AI."

Technical architecture reconstruction: from "step-by-step synthesis" to "one-step synthesis"

On the technical level, AssetGen 2.0 uses a completely different architecture from its predecessor:

• Version 1.0
  This is done in two steps: first, generate multi-angle 2D images of the target asset based on the text prompts, and then synthesize the 2D images into a 3D model through another neural network;
• Version 2.0
  Upgraded to a single-stage 3D diffusion model, 3D assets are generated directly from text prompts, and the training data comes from a massive native 3D asset library.
  This "one-step" generation method improves the model's ability to control spatial structure and detailed textures from the underlying logic.

Promoting the democratization of creation: the ambition from "tools" to "ecology"

AssetGen 2.0 is currently being used within Meta for 3D world development and is expected to be available to Horizon desktop editor users later this year.

Officials said the technology will "promote the popularization of 3D content creation, making it as easy to use as 2D content creation, and opening up new creative possibilities for artists, designers, and developers working on the Horizon and Avatar platforms."

It is worth noting that Meta also revealed that it is developing "complete 3D scene generation" AI - in the future, creators will no longer need to generate individual assets one by one, but can generate a complete virtual world in minutes based on text or image prompts.

This means that AI-assisted digital content production is moving from “parts manufacturing” to a new era of “entire city construction.”

As generative AI expands from text and image fields to 3D space, the implementation of AssetGen 2.0 will not only reshape the content ecology of the virtual social platform Horizon Worlds, but will also likely drive an efficiency revolution in industries that rely on 3D modeling, such as e-commerce virtual display, game scene construction, and architectural visualization.

【Terminology Analysis】

3D Diffusion Model : A generative model that directly generates three-dimensional data. It achieves high-quality content generation by gradually reducing noise in three-dimensional space. Compared with traditional 2D to 3D solutions, it can more accurately capture the spatial structural characteristics of objects.