AVATAR-AGENT: A Multi-Agent System for Expressive 3D Avatar Generation
Year: 2026
Type: conference-paper
Abstract: While recent advances in generative AI have enabled text-to-3D avatar synthesis, platforms built on creator-made content ecosystems require approaches that guarantee asset compatibility, maintain established visual styles, and preserve creator contributions. Digital avatars in platforms like Roblox are assembled from modular 3D components (assets) such as body parts, clothing, and accessories. Rather than synthesizing new holistic avatars or 3D assets through generative models, this paper introduces a fundamentally different paradigm: agentic retrieval-based avatar generation. We present AVATAR-AGENT, a multi-agent system that applies the planning, memory, and self-correction capabilities of Large Language Models to intelligently compose avatars by selecting and assembling components from a discrete corpus of existing creator-made assets. This approach grounds generation in curated inventory, ensures visual consistency, and supports creator economies while providing natural language control. The system operates through a four-phase hierarchical pipeline: intent understanding, strategic planning, execution and assembly, and verification with iterative refinement. In evaluation, AVATAR-AGENT achieves 94.0% prompt constraint satisfaction with quality scores statistically comparable to avatars created by humans, while substantially outperforming the semantic retrieval baseline used in Roblox’s production search system (72.6% constraint satisfaction). By reframing avatar generation as an agentic retrieval-and-assembly problem rather than a generative modeling task, AVATAR-AGENT demonstrates how LLM-based agents can enable intuitive creation workflows while respecting ecosystem constraints and creator contributions.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7162017592
