Agenda

This presentation introduces IntelliScene, a multi-agent 3D scene placement system for game production. Its core enables AI to complete scene layouts via "reasoning" instead of fragile handwritten rules. It encapsulates senior artist experience into data, allowing agents to conduct "slow thinking"—a form of Deliberative Reasoning. Linguistic agents use Chain-of-Thought (CoT) to parse requirements, generate a scene graph, and auto-invoke geometric solvers.The speech will explain the evolution from the LLM-based 1.0 to 2.0, which incorporates visual guidance (style-aligned image models, VLM parsing, asset retrieval, and 6D pose estimation). It will cover method iterations, results, and implementing multi-agent AI into existing art pipelines.

Takeaway

Attendees will learn to design multi-agent reasoning pipelines for 3D scene placement, integrating LLM deliberate thinking, visual guidance, geometric solving, and physical constraints. They will acquire data strategies, system models, and practical experience transferable to their own tools and environmental art workflows.

Intended Audience

Intended for practitioners (tech art leads, programmers, AI engineers, artists) applying multi-agent and vision-language models to 3D scene auto-placement. Attendees should be familiar with game engines and have a basic understanding of deep learning/AI concepts.