How Microsoft’s Open‑Source TRELLIS.2 Generates Full‑Texture 3D Assets in 3 Seconds

Generative AI has achieved large‑scale success in 2D domains such as images, video, and text, but 3D generation remains a difficult frontier because it tests representation, learning objectives, and engineering usability simultaneously.

The core difficulty is not merely producing a visually plausible object; models must preserve geometric consistency, semantic stability, and structural usability across high‑dimensional space. A model that looks correct from a single view may collapse when the viewpoint changes, or it may generate photorealistic surfaces that cannot be exported as editable 3D assets.

Recent research has explored several technical paths. NeRF‑based methods excel at visual continuity but are inherently rendering‑oriented and struggle to provide meshes, topology, or physical properties. Voxel or explicit‑mesh generators produce clear structures yet suffer from limited resolution, detail, and generalization. Single‑view or few‑view approaches improve efficiency but often lack multi‑view consistency and stable geometry.

These divergent attempts reveal a deeper systemic mismatch: the representation, generation pathway, and training objectives are misaligned. When a model’s optimization target focuses on “looking reasonable” rather than being “structurally sound,” the results cannot bridge the gap from demonstration to practical application.

To address this, Microsoft Research Asia released TRELLIS.2, which introduces a novel field‑free sparse voxel structure called O‑Voxel. This representation enables the generation of arbitrary‑topology objects with rich material attributes (metal, plastic, glass, wood, water ripples) and fully constructs internal geometry. TRELLIS.2 also achieves 16× spatial compression, allowing a 40 billion‑parameter model to train and infer efficiently. In practice, it can generate a 512³ full‑texture 3D asset in roughly three seconds.

The TRELLIS.2 3D generation demo is hosted on the HyperAI website under the “Tutorial” section (https://go.hyper.ai/1nofM). Users can run the demo by following these steps:

Visit the HyperAI homepage, select “TRELLIS.2 3D Generation Demo,” and click “Run this tutorial online.”

Clone the tutorial repository into your own container via the “Clone” button.

Choose the “NVIDIA RTX 5090” hardware option and the “PyTorch” image, then select a pricing plan (Pay‑As‑You‑Go, Daily, Weekly, or Monthly) and click “Continue job execution.”

Wait for the resource allocation; once the status changes to “Running,” click “Open Workspace” to enter the Jupyter environment.

Inside the workspace, open the README, click the top‑right “Run” button, and monitor the execution. When completed, use the provided API link to view the demo results.

HyperAI offers new users a promotional credit: for just $1, users receive 20 hours of RTX 5090 compute (regular price $7), which remains valid permanently.

The tutorial concludes with a link to the demo repository (https://go.hyper.ai/1nofM) and invites readers to experience the high‑efficiency 3D generation capabilities of TRELLIS.2.