Meta has revealed a new version of its 3D asset generation AI system that it says is a "significant leap forward".
The Meta Horizon Desktop Editor tool, used to create Horizon Worlds experiences, added the ability to AI-generate 3D assets from text prompts last month, leveraging Meta's original AssetGen model from 2024. This lets anyone build a custom virtual world in a matter of hours for free, with no 3D modeling skills required. But these AI-generated assets have noticeable quality issues, especially visible close up in VR, and that's what AssetGen 2.0 is meant to address.
David Heaney
AssetGen 2.0 delivers "dramatically improved detail and fidelity" compared to the original, Meta claims, including "geometric consistency with very fine details". The company says AssetGen 2.0 sets a new standard for the industry, and "pushes the boundaries of what's possible" with generative AI.
On a technical level, AssetGen 2.0 has a completely different architecture than the original. AssetGen 1.0 worked by first generating multiple 2D image views of the intended asset, based on the prompt, then feeding these images to another neural network which generates the asset. In contrast, AssetGen 2.0 is a single-stage 3D diffusion model directly generating the asset from the text prompt, and was directly trained on a large corpus of 3D assets.
3D assets generated by Meta's AssetGen 2.0 AI model.
Meta says it's already using AssetGen 2.0 internally for 3D worlds creation, and will roll it out to the Horizon Desktop Editor later this year.
"We expect AssetGen 2.0 to help democratize the creation of 3D experiences, making it as accessible as 2D content creation and opening up new creative possibilities for artists, designers, and developers working on Meta’s Horizon and Avatar platforms", the company writes.
Meta also teased that it's working on AI generation of "entire 3D scenes", which will enable creators to generate entire worlds using text or image prompts, not just individual assets.
