This is very cool. Fast image generation is nice, but imagine you can send 24x24x2 bytes = 1152 bytes and you can share a photo.
Applications: super low bandwidth projects such as radio transmission. Send a map or a photo of a location. Or a family photo.
For example with the settings on the picture, you can realistically send the image in 8 seconds (plus encoding and decoding on the phone).
Imagine that you could do the same with audio messages with sound ai models. I think there are very interesting data compression implications of these advanced statistical models.
Considering a duty cycle is also important. For Lora to work, a 1% on air transmission is enforced, so if you broadcast for 8 seconds, you should be silent for 792 seconds. Of course you can have different LoRa parameters that make it even more feasible. 9Kbps is feasible. You can play with your Lora parameters here:
https://unsigned.io/articles/2022_05_05_understanding-lora-parameters.html
Stable Cascade here:
https://github.com/Stability-AI/StableCascade
#m=image%2Fjpeg&dim=1074x992&blurhash=%3B25%3D9G%7EX-p-%3D.8M%7BIAM%7BM%7Bt7kCt7f6WCt7WBRjj%5BoLWBM%7BayWBM%7BRjofofayWBWBWVoJWAa%7DofjtRPRjkCWAoKt7ozofofo0WBNGofoLWBayofofRjWBfQayWBofofayayRkj%40t7WBbHt7ofayWB&x=cd7e895e45592c6b6d57bd9b57d9ec7563b2c4d307b8b1af6d36270808dd0821
That's really interesting, there is probably entirely new field of compression possibilities that haven't been tapped yet.
This (what you describe here) is an extreme case, but I'm sure there are usecases for the same principles at higher rates. Once there is some kind of ubiquitous hw acceleration for crunching AI it will probably be used in most things.
It's not very new, the encoder-decoder architecture has been used for a long time now, so with neural networks you get compressed representation of data as a byproduct quite often.
