Agenda

This session introduces Zenith, a hybrid pipeline that automates the creation of multi-layered top-down maps from 3D environments. By combining procedural geometry extraction with a multi-encoder ControlNet architecture, Zenith generates walkable areas and stylized layers—line art, shadows, and highlights—using depth, normals, and detail conditioning. By unifying multiple diffusion models, Zenith delivers visually consistent results without requiring large training datasets. We will explore how it enhances efficiency, share insights from artist feedback, and discuss practical lessons for integrating diffusion models into game development.

Takeaway

Attendees will explore how Zenith streamlines the creation of stylized top-down maps from 3D environments by combining procedural geometry processing with a custom multi-encoder diffusion model architecture. The session will detail how Zenith generates walkable areas and artistic layers—such as line art, shadows, and highlights—using fine-tuned ControlNet models, while preserving visual consistency and reducing manual effort. Technical participants will gain insights into training strategies, multi-modal conditioning, and architectural design choices that enable high-quality outputs with limited data. The talk will also share practical lessons from integrating Zenith into production workflows, emphasizing how automation can enhance creative efficiency without compromising artistic control.

Intended Audience

Technical artists, tool developers, software engineers, graphics engineers, applied scientists, machine learning engineers