Generative AI Manipulation for 3D Texture Synthesis

Generative AI models have revolutionized 2D image tasks, including manipulation, style transfer, and novel image synthesis, enabling neural networks to create complex and aesthetically appealing visuals. However, similar progress in 3D object generation has been hindered, partly due to the scarcity of large, diverse datasets. In response, we present Deep Single-Object Manipulation (DeepSOM), a novel 3D manipulation framework based on a convolutional backbone. Our approach requires only a single input pair: a source object for manipulation and a primitive proxy. These inputs are augmented through random free-form deformations to generate a diverse training dataset. The model learns to map deformations of the primitive to corresponding changes in the complex object, allowing even minor perturbations to produce highly variant shapes. We evaluate DeepSOM's effectiveness in "freehand" synthesis of spatially textured surfaces, demonstrating its potential for both graphical renderings and the creation of physical objects.