Idea for a next gen image compression method:

* Put your image into a reverse AI image generator to get the prompt for that image.
* store the text string as the compressed image
* Put the text string into to AI image generator to get back your compressed image.

https://petapixel.com/2021/08/...

Lossy downscale plus ai upscale plus optional error correction is probably already here

@atheist you don’t need upscaling/downscaling if all the info is contained in the prompt text.

People are doing the same thing for email.

AI please write a lengthy email given these two sentences.

AI please condense this lengthy email down to two sentences.

@BobbyTables that’s the opposite, they’re doing inflation!

You can't predict the outcome, the same prompt will give you a different image every time. Suddenly, the cover of Bad Boys 4 is two white dudes holding a carrot because there was a new wave of protests involving race and someone wrote on reddit that carrots are the new weapon against obesety. AI only read the first part.

@devdiddydog yes. You can see it as the "lossy" part of lossy compression. 🙃

Encoder-Decoder models and google were already doinge this 15 years ago. Except instead of encoding into a human readable prompt you leave it as the compressed weights.

@Hazarth nice. How does the size of those weights compare to the size of the human readable prompt?

@Lensflare much worse, obviously, but the recreation is actually precise, while a prompt will generate a $max_seed unique images with the same conceptual meaning, the compressed encoding will actually recreate it exactly (It's lossy compression, but lateste attempts outperform jpeg2000)

You can see some examples at the bottom here: http://lherranz.org/2022/08/...

@Hazarth that’s interesting. I wonder if that’s how the human brain stores and recalls images.

@Hazarth if it even remotely resembles how real brains work, it could be an explanation or a hint why memory transfer is not possible.
If you’d transfer a memory to a brain which has a different model, then obviously the weights wouldn’t match and the retrieved memory would be complete noise.

@Lensflare hmm, but I think you could train a middleware that translates one memory to another if you have an encoder and a decoder for two people then it should be connectible. But then again, human menory is a lot more complex since It's not Just image data :D

@Hazarth and he delivers! Thanks Hazarth.

you could use llava for the prompt and then stable diffusion for the image