I initially built this tool with all Python libraries packaged inside of the add-on using .WHL files for each operating system and versions of operating systems. However it got incredibly complex managing over 100+ wheel files across Python packages and the size of the actual addon was getting way too big to manage. Just the Pytorch library alone only on Linux was over 1 GB, and that's just for one library. In summary, it wasn't feasible to try to manage 100+ .WHL files across operating systems and machines if I wanted to scale and develop the tool further. That's why the libraries aren't packaged inside of the add-on. And that decision has allowed me to keep the core code of the add-on lean and lightweight to install and update with new versions. Otherwise it's a pain not only for me, but also users as well.

I'm definitely not planning to use the Midas model as the "go-to" depth model in the future, however it's a great foundation and starting point while developing the tool. It might not be the most advanced, but it's stable and gives me a place to start with to build the tool.

No. The necessary Python libraries to use the tool to generate depth maps is only installed after the user decides to either click on the "Install All Packages" button the the Preferences window, or the individual "Install" buttons for each library.

In the Preferences window it displays all the Python libraries that will be installed after the user chooses to do so. Also, there's a little info icon for each library in the Preferences window that you can click on that will take you to the official page for each Python library. The libraries that will be installed include cffi, cycler, dateutil, facenet-pytorch, joblib, kiwisolver, matplotlib, mpmath, opencv, packaging, pillow, psutil, pycparser, pyparsing, pyyaml, scipy, six, sklearn, sympy, threadpoolctl, timm, torch, torchvision, tqdm, typing_extensions, and ultralytics.

The MiDaS model weights file contains the pre-trained parameters used by the MiDaS deep learning model to predict depth from 2D images. These weights are learned from large datasets during training and allow the model to understand depth-related patterns, such as distances between objects in an image. When generating depth maps, the model uses these weights to accurately infer the relative depth of various objects, creating a map where closer areas appear brighter and farther areas darker. This enables realistic 3D-like representations from flat images, which is essential for many visual effects and 3D reconstruction workflows.