OpenAI reveals a brave new AI accomplishment with Point-E
This 3D model was trained on tagged photos, same as OpenAI’s Dall-E 2 and Stable Diffusion
OpenAI released Point-E to the public, a machine learning system that creates a 3D object from a text prompt. Point-E can make 3D models in one to two minutes on a single Nvidia V100 GPU.
The name ‘Point’ refers to point clouds which are a discrete group of data points in space that reflect 3D shapes. The ‘E’ in the name refers to efficiency as it promises to be faster than other methods of making 3D objects.
Point clouds are easier to make computationally but they currently cannot capture an object’s fine-grained structure or texture. To get over this limitation, the Point-E team trained a second AI system to convert the clouds at Point point E to meshes, meshes are collections of vertices, edges and faces and are widely used to define things in 3D modelling.
In addition to the mesh-generating model, which is an independent model, Point-E consists of two additional models: a text-to-image model and an image-to-3D model. To understand the connections between words and visual concepts, the text-to-image model was trained on tagged photos.
OpenAI researchers claim that the Point-E point clouds might be used to fabricate real objects, such as through 3D printing. The next significant step could be model-synthesising AI. The fields of science, architecture, interior design, film, and television regularly use 3D models. Models are used by engineers to create new machinery, vehicles, and structures.