IDEAVATAR: Identity-Preserving Avatar Generation with Controllable Emotions

3D human generation has made notable progress and seen broad adoption in areas like virtual avatars and digital asset creation. Yet, existing methods still struggle with jointly modeling facial expression control and identity preservation, often failing to balance the semantic flexibility of expression with the structural fidelity of identity, resulting in degraded facial details or inconsistent expressions. We propose a novel 3D human generation framework conditioned on multimodal image and text inputs, enabling precise identity preservation and expressive facial control. Our approach innovatively combines decoupled identity encoding, CLIP-based expression guidance, and a geometry-aligned head-body fusion process, achieving detailed, coherent, and flexible 3D modeling. Our method is compact and effective, offering interpretable modeling, flexible control, and strong generalization across diverse identities and expressions. Experiments on the FaceScape dataset show that it outperforms existing baselines in facial detail, identity consistency, and expression alignment, achieving over 87% facial confidence within 3 minutes of inference, demonstrating strong modeling performance and practical value.