Abstract
This report introduces Make-A-Character 2, an advanced system for generatinghigh-quality 3D characters from single portrait photographs, ideal for gamedevelopment and digital human applications. Make-A-Character 2 builds upon itspredecessor by incorporating several significant improvements for image-basedhead generation. We utilize the IC-Light method to correct non-idealillumination in input photos and apply neural network-based color correction toharmonize skin tones between the photos and game engine renders. We also employthe Hierarchical Representation Network to capture high-frequency facialstructures and conduct adaptive skeleton calibration for accurate andexpressive facial animations. The entire image-to-3D-character generationprocess takes less than 2 minutes. Furthermore, we leverage transformerarchitecture to generate co-speech facial and gesture actions, enablingreal-time conversation with the generated character. These technologies havebeen integrated into our conversational AI avatar products.