Take It Out

An exact generator trained on six public-domain draughtsmen and a pool of museum paintings. Every pixel it makes is an exact blend of training pixels with named owners. Take an artist out and see what changes, and who gets paid.

loading…

1. The picture

the picture with the artists shown above

what changed, times the gain

Every combination of artists (64 per picture) was computed exactly by the engine when the receipt was made; switching them here reads those results. Nothing on this page is estimated or drawn by hand.

2. Who gets paid

use: what the receipt says was copied leave-one-out: what changed when that artist alone was removed Shapley: average change over every order of removal

3. A drawing as a player

A viewer's drawing starts the generator part-way (strength k = 14 of 19 steps, where the drawn shape first shows; at lower strength every drawing gave nearly the same paper-toned picture). You can take out its paper (a blank sheet in the same colour) and its lines (the drawing minus that sheet) separately; every one of the 256 combinations was run. Pay treats the drawing as one player: splitting a party into more players moves everyone else's pay, so the split is shown as description only.

the drawing

the picture with the players shown above

Draw your own

draw here (paper and ink are the corpus's own colours)

the engine's picture appears here

Turning a drawing into a picture runs the generator on a local server, so these two buttons are off on this site.

4. The arithmetic

Shapley value = sum over every set S of the other artists of weight(|S|) × [value with S plus this artist − value with S alone]. weight(s) = s!(5 − s)!/6!. Value of a set = minus the distance from its picture to the picture with every artist.

5. Checks

With this picture

From the frozen tests behind the instrument

6. Receipt

Replaying needs the generator itself, so on this site the button is off: download the receipt and run the check locally. The spot check regenerates 4 of the 64 combinations and checks that their pictures and values match the receipt exactly. It can miss a forged value in a combination it did not pick; the complete check regenerates all 64: python3 receipt/verify.py <receipt> --replay all.

7. What this is and is not