Learning an Interpretable Stylized Subspace for 3D-Aware Animatable Artforms.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Traditional paintings are static, limiting viewer interaction.
• Existing methods lack interpretability and fine-grained 3D control for artistic content.

Purpose of the Study:

• Introduce 3DArtmator, a novel framework for 3D-aware animatable reconstruction and editing of artforms.
• Enable interactive exploration of static masterpieces by creating an interpretable, stylized latent space.

Main Methods:

• A two-stage optimization framework is proposed.
• Stage one identifies an anchor in the original latent space to mimic art content and pose.
• Stage two trains a customized 3D-aware Generative Adversarial Network (GAN) with a style-directional difference loss to preserve latent structure.

Main Results:

• 3DArtmator successfully generates 3D reconstructions and enables editing across diverse art styles.
• The method maintains interpretability and 3D controllability within the stylized latent space.
• Experimental validation confirms the effectiveness and versatility of the proposed approach.

Conclusions:

• 3DArtmator offers a new paradigm for revitalizing artforms through interactive 3D manipulation.
• The framework opens avenues for enhanced artistic engagement and exploration.
• Interpretability and 3D control are key contributions for digital art applications.