AI model generation

Happy New Year everyone

I have ended 2022 playing with the generative modelling tool Point·E from OpenAI. Here are 33 results from the test prompt “a building”

Each building takes about 2 minutes to generate on my little GeForce GTX 1650 GPU or nearly one hour on my 16 core 17 CPU. It then takes 3 minutes to create the surface model from the pointcloud. They all fit within a box of 1000x1000x1000 and have no specific units.

I particularly like this one:

To try it out at the moment you need some python skills : here is how I did it:

  • set up a new python environment (I used venv but it should work with conda)
  • (for cuda on the GPU) Follow the pytorch setup instructions from Start Locally | PyTorch
  • clone the GitHub repo
  • Install the Point-E package
  • Install and launch jupyter notebook
  • Open the example jupyter notebooks and play around

I am open to ideas for other text prompts : what prompt would you like me to try? I find simple ones work best.

4 Likes

Damn, that’s a conservative estimate of what a building is. :yum:

1 Like

Here are the AI generated pointcloud and the resulting mesh for my favourite building - you can see that fine details like the spire are lost by the marching cubes meshing process.


a building_2.ply (211.0 KB)
a building_2.xyz (612.4 KB)

1 Like

12 versions of “a laptop computer”