How Tripo Approaches Clean Topology in Its 3D Asset Pipeline

In this 80 Level interview, Tripo Co-Founder and Chief Scientist Yanpei Cao discusses the company’s approach to making AI-generated 3D assets usable in production. The focus is Smart Mesh and the Nexu

Creating 3D models from images or text with AI has become faster, but the resulting output does not automatically mean it is an asset ready for use in production. Yanfei Chao, co-founder and chief scientist of Tripo, pointed out that messy topology, difficult-to-edit materials, and inconsistent geometry increase the amount of follow-up cleanup. The problem Tripo is trying to solve is closer to “usability” than simply improving generation quality.

The Tripo Smart Mesh he introduced is a path for creating game-ready assets from a single image, based on the SIGGRAPH research project Nexus. The goal is to provide an orderly edge flow instead of a mesh with polygons tangled together chaotically. Chao explained that this approach makes animation work easier and polygon counts simpler to control, reducing the burden of passing assets into tools such as Blender, Maya, and Unity.

The input can reportedly take various forms, including not only text prompts but also a single reference image, multiple images, and character sheets. Tripo is running a path for high-resolution models alongside the Smart Mesh path, and for the latter it has set the goal of producing results with controllable polygon counts and usable edge flow within 10 seconds. The example mentioned in the interview is a character with approximately 20,000 polygons.

Rather than claiming that automation alone can finish every task, the company emphasizes enabling artists to repeatedly revise results and maintain art direction. Follow-up features such as Magic Brush, which adjusts details in geometry and textures, along with part separation, texturing, rigging, and animation, fit into this workflow. Chao predicted that expanding from single-asset generation to environments, interactions, and dynamic worlds will require high-quality asset generation together with logic and code, as well as world models.

Maru