Vehicle Form Generation Algorithm and Application Based on Diffusion Models

This research is inspired by the 2D generation of guided diffusion models and proposes a guided near-vehicle shape point cloud generation method based on diffusion models. It designs a point cloud encoder-decoder based on a fully connected network, a 2D feature grayscale image generation model based on DDPM, and a 2D classification feature grayscale image LoRA model. It uses a dataset integrated from a self-built parametric shape representation set and the ShapeNet dataset for pre-training and finally integrates into the CarStylingDiffusion model. The CarStylingDiffusion model can achieve the transformation from text to the abstract 3D quadrilateral mesh model of a car, greatly improving the quality and efficiency of shape generation. Under the condition of a given dataset, the indirect test shows that the generation effect of the CarStylingDiffusion model is similar to PointFlow, basically meeting the requirements of shape generation. Designers can input the desired style and vehicle information, use the CarStylingDiffusion model for the generation of car abstract shapes, and further use it in the subsequent design process.