Double Exposed Archive
The process it uses is something like this:
- Pick one image at random using some keywords like "person".
- Pick another image at random using some keywords like "animal", "sea", "machine".
- Make grayscale, clean up, and crop to center.
- Combine the images using the "screen" blending mode.
- Apply a color gradient.
While I'm quite happy with the results, one problem area is that some kinds of line art or blurry images just don't work well with the bot, as they tend to obscure the other image. At the moment I'm looking into training a neural net to recognize "bad images" and prune them out; we'll see how that goes. Ψ