Towards ML-based Assessment of Synthetic Characters Heads
Igor Borovikov, Karine Levonyan, Panda Elliott, Étienne Danvoye · 2025
Virtual environments present ever-growing requirements for their population of synthetic characters. In many applications, various character heads must provide a balanced representation of age, gender, and ethnicity. With a character count well above 10,000, manually checking and verifying the target metrics is impractical. This paper outlines a possible pipeline for generating parametric avatar heads. The main focus is the final stage, where generated character heads are evaluated for aesthetic quality metrics. The proposed quality assurance (QA) approach uses ML models trained on sparse data obtained from human evaluation. The QA ML models’ training data collection leverages in-house crowdsourcing and aims to match the assessment initially provided by the expert art direction. We illustrate the approach with heads generated using FLAME.