PartCrafter is a methodology for creating structured 3D networks based on combining latent diffusion models and transformers capable of decomposing an object into its components.
A recent similar development from ByteDance has been released, which I recently mentioned — it also works with input images and produces a complex 3D model of an object.
The developers claim that their approach does not involve segmenting the original image, but relies solely on machine learning to recognize parts of the model directly in the latent space. For this, they conducted extensive training on very large datasets.
There are no demo versions available yet, and the source code is planned to be released no later than July 15.