HFGen: 3D Pose Generation and Emotional Tone Control for Historical Figures

HFGen is a 3D pose generation and emotional tone control model specifically designed for historical figures, aiming to enhance interactive experiences in history education through advanced deep learning techniques. Existing methods often fall short in generating natural 3D poses and expressive emotions, typically relying on manual guidance or skeletal image manipulation. These approaches lack smooth linguistic transitions and fail to capture emotional nuances embedded in historical datasets. HFGen addresses these challenges through a three-stage pipeline: standard character creation, animatable avatar learning, and emotional tone generation. First, in the standard character creation stage, a trainable Neural Radiance Field (NeRF) is initialized with SMPL priors. To improve geometric accuracy and accelerate convergence, we employ 3D-consistent Score Distillation Sampling (SDS), enhancing supervision consistency. Next, in the animatable avatar learning stage, human poses are randomly sampled using VPoser from the SMPL model. A Density Weighting Network is introduced to improve the animation of complex geometries, enabling accurate simulation of arbitrary motion sequences. Finally, in the emotional tone generation stage, the emotional space of a large language model is fine-tuned to align generated text with the emotional expressions found in the input dataset. Experimental results demonstrate that HFGen performs exceptionally well in both 3D character generation and emotional text processing, offering a powerful technical solution for the digital reconstruction of historical figures.