on Toshareproject.it - curated by Bruce Sterling
*It shows the principles.
Criteria: plausibility, proportionality, and originality
Ideally, we want an image generator to exhibit:
Plausibility: individual generated images should be indistinguishable from “real” images (no weird artifacts or glitches).
Proportionality: a bunch of randomly-generated images should be indistinguishable from a bunch of randomly-selected real images (no traits should be noticeably over- or underrepresented).
Originality: Image generators can train by looking at real images, but image generators should not just copy, palette-swap, splice-together, or otherwise misappropriate these images. A bunch of randomly-generated images should be indistinguishable from a bunch of randomly-selected real images even if we plagiarism-check both sets against the training dataset….