Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

For DALL-E and Stable Diffusion, the model size is an order of magnitude smaller than the total size of all the training set images? So it's not possible for the model to regurgitate every image in the training set exactly?

For Copilot, is there a similar argument? Or its model is large enough to contain the training set verbatim?



Mentioning DALL-E, you've hit on something.

The world seems slightly mad about these things that produce "almost" pictures from text. We forgive DALL-E when it produces a twisted eye or an impossible perspective, because its result is "close enough" that we recognise something and grant the image intention.

So now you've got me waiting for DALL-Ecode. Give DALL-Ecode a description, it produces code.

"DALL-Ecode: Code that is sufficiently close to what you'd expect that you'll try to use it."

"DALL-Ecode: Code that looks like it does what is needed."

"DALL-Ecode: Good enough to compile, good enough to get through a code review (just not good enough to get through testing)."


How small is DALL-E/SD, compared to say, training dataset images shrank to 120x120, JPEG compressed at q=0.3, compressed as .tar.bz2?


> The data can comfortably be downloaded with img2dataset (240TB in 384, 80TB in 224)

https://laion.ai/blog/laion-5b/

Not exactly what you asked, but hopefully useful? The model weights are about 4 GiB I believe.


IIRC, 2.5 billion images were used to create a 4.5GB dataset. That is less than 2 bytes per original image.




Consider applying for YC's Fall 2025 batch! Applications are open till Aug 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: