PerSynth: Personalized Character Video Synthesis from Text Prompts

Recent human image animation methods predominantly rely on motion signals extracted from driving videos, which limits the diversity and flexibility of generated actions. To overcome these constraints, we propose a novel pipeline that synthesizes human motion videos from a single reference image guided by free-form text descriptions. Our approach first generates 3D pose sequences from text using a dedicated text-to-motion model. These poses are then reformatted into a standardized key-point representation compatible with existing animation engines and aligned to the reference image’s pose to preserve identity and spatial coherence. Finally, we leverage pretrained image-to-video synthesis models to produce high-quality, temporally consistent animations without additional retraining. Preliminary results demonstrate that our method can render diverse, text-driven motions while maintaining visual fidelity and structural consistency. For more details, code and additional examples, please visit our project page: https://bmhungqb.github.io/PerSynth