Image-to-Image Translation Demo | Hacker News Reader

Image-to-Image Translation Demo

affinelayer.com

122 pointsma2rten10 comments

Looks very interesting (and slightly scary) but does seem to a bit under the wheather right now.

I drew a boot, got a cat [1]. That's some serious over-fitting, or something. I hope no cats were hurt.

[1] Perhaps this works: http://imgur.com/a/xvg6q.

That's because that was run on the cat's dataset. For the boot use the boot editor?

Huh? I did, it says "edges2shoes" right there in my image, and I cleared away the example shoe before drawing my crappy attempt at a boot.

Of the four editors on the page ("facades", "edges2cats", "edges2shoes" and "edges2handbags") I clearly used the one related to shoes/boots.

The backend service was probably confused. Just try running it again.

It says "Chrome is recommended" so I thought it was running in browser, but then when trying it I got "error connecting to server". I'm guessing it's getting the hug of death and is unusable just because it's trying to upload some telemetry or something. Please make it so telemetry doesn't fail the demo...

> The interactive demo is made in javascript using the Canvas API and talks to a backend server that runs the images through Tensorflow.

It's because it's server side, it was using a Google hosted Tensorflow service that I disabled since it was breaking things. The amount of telemetry that is not on here would blow your mind.

I got this vision of horror based on a poorly drawn snail: https://i.imgur.com/N1UBn2G.png

The fact that the source didn't include any active pixels at the top (compared to the result) seems a bit weird though.

Here kitty kitty http://i.imgur.com/Ua4wx8s.png

Makes sense to me as I figure all the training data had colored pixels all the way to the top and that some recognition of cat parts would include an edge of the frame and not always be perfectly centered.

Reply on news.ycombinator.com