Give users a choice - do they want to draw a frame (will take a couple of minutes, many people don't have time/patience) - or do they want to vote on a few frames?

For each frame, get at least 4 different drawings.

For voters, show them 2-4 frames and have them pick the best one.

If there are more than 4 versions of a frame, do a bracket-based competition to determine the best frame.

This could be an ongoing process. Generate a version of the movie using the best available frames, but continue allowing people to redraw frames. The movie would continue getting better and better over time.

Then, when you watch the movie, maybe you could have a slider to control which "generation" of the movie you're watching. Slide all the way to the left and it will be really primitive, slide all the way to the right and it will be the best version.

I can't even begin to guess how you would numerically average or compare hairline drawings.

If they were fully shaded frames it would work, but hairlines I have no idea.

Also I observed that some frames left a dark background white, and others filled in a dark background as black. So you'd have to set up a lot of drawing "rules" to ensure it was even conceptually meaningful to generate an average in the first place.

Show the voter 4 drawings.

They pick the worst one.

Throw in a control once in a awhile.

Eventually you’ll have a sorted list of the best to the worst frames. Have the worst ones redrawn.

However, if you let it run too long, you might get a Pixar film.

Fun alternative: show artists the best images. Eventually it will be gamed and it will become the Bee Penis film.

Flag images which have less than a certain % covered. Probably garbage.

Pass the reference frame and each candidate drawing through a CNN. Measure the cosine similarity between the candidate embedding and the reference embedding. Flag drawings that fall below a threshold.

Pass candidates through CLIP. Flag images with obvious garbage descriptions.

Use distance fields.

The problem isn't just approximate positioning of lines, but radically different artistic styles and amounts of shading and detail. Is someone's hair made of one line, five lines, or 100 lines? Does a face have 3 simple lines or 50 little lines that give a sense of skin texture?

Distance fields would work great for font outlines, but not for translating movie frames.

Oh, I don't know. How well does it have to work? If anything I'm suddenly very curious what kind of results you'd get from converting line art to distance fields, averaging/interpolating, and converting back. Maybe I'll write some code...

Throw out whole frames that have a large enough Wasserstein distance?