We've modified the input scripts and are starting on our journey of training the models:
https://github.com/webaverse/GET3D/blob/master/render_shapenet_data/render_shapenet.py
We got much faster and more game-like generation results by switching to Eevee renderer and going to HDRI. Eevee renders 24-frame multiview in ~3-4 seconds per model, and it can crank through an entire average-sized shape category set in about an hour or two.
Also, because it's a realtime-style rasterizing renderer, it's pixel perfect and not subject to noise. To compensate for a lack of dimension we've lit the scene with a studio HDRI and applied an ambient occlusion post processing effect.
So, questions for the team:
- do you see any issues with this, is there a requirement for the realistic renderer?
- would you accept a pull request for this? we will need to modify our code since we are also doing some transformation to normalize the ShapeNet V2 dataset, but I could make a PR to switch between eevee and cycles.
We've modified the input scripts and are starting on our journey of training the models:
https://github.com/webaverse/GET3D/blob/master/render_shapenet_data/render_shapenet.py
We got much faster and more game-like generation results by switching to Eevee renderer and going to HDRI. Eevee renders 24-frame multiview in ~3-4 seconds per model, and it can crank through an entire average-sized shape category set in about an hour or two.
Also, because it's a realtime-style rasterizing renderer, it's pixel perfect and not subject to noise. To compensate for a lack of dimension we've lit the scene with a studio HDRI and applied an ambient occlusion post processing effect.
So, questions for the team: