Hi, thanks for open sourcing this repo, it's really useful for rendering compositional scenes!

After reading through the code, I found that there were some placeholders for objaverse, I guess there is a plan to support fetching objaverse samples on the fly and render a compositional scene.

I have implemented this feature on my fork, as follows:

1. I use this filtered id list (filter some bad prompts and missing textures) of objaverse: https://github.com/ashawkey/objaverse_filter/blob/main/kiuisobj_v1_merged_80K.csv
2. Sample an id from the csv, and download from https://huggingface.co/datasets/allenai/objaverse by concatenating the instance id to the base url. Notice that here the objaverse-v1 instead of objaverse-xl is used since the latter has more low quality samples, I think
3. After rendering, the downloaded temp objects will be cleanup and remove
4. An example yaml is provided

Besides, I have added some postprocessing logics when importing a glb into the scene in blender_utils.py::add_object_file, since sometimes imported glb is not clean (e.g. some examples in objaverse, it may contains a number of empty objects with transforms), which will make the rescale / rotation transformation fail during rendering

Some example renderings:

Hope this can help, but feel free to modify this PR or close it if it's not planed or intended :)