Citation: Discriminative Unsupervised Feature Learning with Exemplar Convolutional Neural Networks (https://arxiv.org/abs/1406.6909)

50,000 MetArt-style nudes. No banners, no ads, no hassles, no distractions. Smart image zoom to fit the woman to the screen.

See random pics (WANDER). When you like one, see the photoshoot on repeat (TRANCE).

To search for something-- e.g., grass, beach, face, pussy-- long press the image (DREAM). This finds nearest neighbors in deep feature space, but only works reliably for simple concepts because it's fully unsupervised. The features you selected by long pressing (inside the box that appears) are more likely, but only very simple searches are "pure".

For example, long press a close-up pussy to see more pussies... find one you like, then press TRANCE to see the rest of her. Long press sand and water to see more girls on the beach. Long press forest greenery to see more girls in the forest. Long press a close-up face to see similar faces... And so on.

If this is not familiar to you, take a look at: http://cs.stanford.edu/people/karpathy/cnnembed/ Consider one of the big "maps" on that webpage. Notice how the tiny image patches clustered together in a map tend to be similar to each other. When you long press the image in Melondream, a box appears. That box is like one of the tiny image patches. Melondream's DREAM shows you images having patches near the patch you selected, in Melondream's map. Also notice how impure even a supervised dream would be.