Ask HN: DALL-E was trained on watermarked stock images?

Re:Ask HN: DALL-E was trained on watermarked stock images? by yreg on Wednesday August 24, 2022 @03:28AM (#32575958)

Op constructed a horrible prompt. First of all, using king Philippe I. is against the ToS, so let's go with a generic "king".

Let's not confuse the AI with "buts", just say that he is giving the speech to cucumbers.

Lastly, specify some style, because this would probably not work out as a photo.

My single try is not bad at all and it could definitely be improved.

https://labs.openai.com/s/3OUmUxKefJCeLhAk4hkeKX4V

[ Parent | on HN ]

Re:Ask HN: DALL-E was trained on watermarked stock images? by yreg on Wednesday August 24, 2022 @04:25AM (#32576393)

I tried it with Stable Diffusion as well. You can use actual people and the model is even pretty decent at many of the famous ones.

On the other hand, it is more difficult to get it to produce absurd results like these.

my prompt: King Philippe I. of Belgium giving a speech surrounded by [[[[large green vertical cucumbers]]]], digital art in the style of Greg Rutkowski

https://files.catbox.moe/1ej1a4.png

[ Parent | on HN ]

Re:Ask HN: DALL-E was trained on watermarked stock images? by throwaway675309 on Wednesday August 24, 2022 @03:24PM (#32584488)

Is the usage of the surrounding brackets some kind of keyword weights specific to stable diffusion?

[ Parent | on HN ]

Re:Ask HN: DALL-E was trained on watermarked stock images? by yreg on Thursday August 25, 2022 @04:30AM (#32591122)

I think it's specific to SD. [Square brackets] increase the weight while (simple brackets) decrease the weight. In the cucumber case I used them to force the model to take into account the less believable part of the prompt, because otherwise stable diffusion often ignores such parts.

[ Parent | on HN ]