Show HN: I made a pictionary game with Stable Diffusion | Hacker News Reader

Show HN: I made a pictionary game with Stable Diffusion

share.streamlit.io

16 pointskn8119813 comments

I used lixica.art to get image-prompt pairs. You are asked to guess the prompt. And given a score based on how close you were to the actual prompt.

Open articleView on HN

CLIP Interrogator Simulator 2022 :)

To be honest I find it pretty impossible to solve.

Ha, I have a running bet with a friend that it is inevitable that there is going to be model soon that is trained to do the inverse, predict the prompt from an image.

There’s a very rich labelled dataset out there now.

Well there is this, but it isn't great yet.

https://colab.research.google.com/github/pharmapsychotic/cli...

All the data is there, you just need a cool half million dollars or so in compute to do the training.

Spun up another instance, in case the streamlit app goes down again, here: <https://diffaki.knhash.in>

Also a correction in the description, the source should read lexica.art

The shorter the prompt the higher the score I get. I got a picture of some anime woman on the beach in a swimsuit.

"anime woman on beach in swimsuit": 0.5

"anime woman on beach": 0.6

"woman on beach": 0.7

"beach": 0.8

"b": 0.815

"": 0.866

> This app has gone over its resource limits.

Hug of death?

Looks like :) Just rebooted the app, looking into more long term solutions.

Link is broken for me now. Says site cannot be reached

Please check again, it should be working now.

Still the same for me. I think it has received a big hug.

I am desperately trying to get in touch with someone to increase my resource limits. Doing a release on a Sunday came back to bite me after all, apologies for the nuisance everyone.

It should be up _for now_.

For those who can't load it, this is how it looks after you enter your guess.

https://imgur.com/nc3XhJn

Reply on news.ycombinator.com