In recent years Blender has gained popularity in the computer vision research community for quickly generating artificial training data for machine learning models. These approaches utilize Blender's Python API to automatically generate images of different poses / lighting conditions and the like.

One of the issues, to my best knowledge, is the limited access to Blender's internal image pixels. We use the compositor approach to generate images, object id masks and the like. While it is (to some extend) possible to save those images using file-output mechanisms it would be great to gain direct access to internal pixel state. Unfortunately only tedious and slow methods (use Viewer Node) are available.

I wonder what would need to be patched in Blender's source to expose internal pixel state directly to Python users.

Best,
Christoph