I couldn't find an automated/artificial intelligence car interior designer, so I decided to experiment and try using generative adversarial neural networks.
For my supercar 3d design concept "Aria Dio", I had already modeled the interior manually in Rhino3d 7, but I was interested to see how well ai would tackle the task. I am reporting these early results of my initial trial in this github repository.
I expect more training time to generate higher resolution interiors, instead of the current somewhat hyper-artlike result.
Use case: This could eventually guide designers along, to create fresh/new interiors.
Car interior imagined after 0 epochs:
Car interior imagined after 200 epochs:
Car interior imagined after 2000 epochs:
Starting to look like an interior, car interior imagined after 9000 epochs :)
Given dataset of "fake" car interiors, as well as real car interiors, a generative adversarial neural network is trained for 9000 epochs to reasonably learn what real car interiors look like, by learning to construct them from this dataset pair of fakes/reals.
-
Collect a dataset of real interiors.
-
Interiors downloaded using python api, bing image downloader.
-
If you want to make a dataset of your own, and you run into issues with the original usage instructions, use God's simple fix; "god_invoke_image_downloader.py", which just does what another author suggested in the issues section of the same page.
-
-
Create a dataset of "fake" car interiors.
- If you want to make a dataset of your own, use my modified version of a sketchify utility, "god_batch_sketchify_utility.py", which I converted to do batch conversion from a directory of color pics to sketched versions, based on this author's code that did the same thing but for single images.
-
Train a GAN on this dataset.
- If you want to train your own, see my quick instructions.
Whether this workflow is the most optimal path is unknown.