|
|
|
|
|
by gd3kr
548 days ago
|
|
Understandable. For context, the GPT in the name comes from an earlier version of this project (https://github.com/gd3kr/blendergpt) which actually used GPT-4 to write python scripts that Blender would then execute. This would allow GPT-4 to program operations like instantiating primitives with the Blender Python API given only a text prompt (ex. "create 50 cubes") The new version of BlenderGPT (lets call this v2) doesn't use an any autoregressive token prediction for the actual mesh generation part, so I understand why it sounds dishonest. I really just chose to stick with the name because artists really didn't seem to care about how the meshes are generated, and the term GPT became closely associated with AI. As for the technical stuff, I've been working on BlenderGPT v2 for the past several months, and until a week ago, i had been using a custom pipeline I built borrowing and re-implementing bits of Unique3D (https://wukailu.github.io/Unique3D/) and combining it with optimized models (flow matching diffusion models etc) for intermediate steps (text to image generation). My optimizations reduced inference time from >2 minutes to only about 20 seconds. This is the model used in this demo i shared: https://x.com/gd3kr/status/1853645054721606100 And then Microsoft released Trellis (https://github.com/microsoft/TRELLIS), and it seemed to leapfrog my model's capabilities on most things. Integrating it into the pipeline wasn't too hard and so I went forward with it. All of this is just to say that there really was a lot of effort put into the core pipeline, and the landing page was mostly an afterthought. Actively working on a more comprehensive one that covers all the points I talked about. |
|