Moving pixels is expensive.
Making them is getting cheap.
JAI is an image format for the world that follows: send the few hundred bytes that say what a picture shows, and let a generative model on the other end draw it.
A description is compression
Every image format trades bytes for distortion. PNG keeps every pixel. JPEG discards detail your eye will not miss. JAI goes to the far end of the same curve: it keeps only what a person would say about the picture, and discards the rest.
That is not a different kind of thing from image compression; it is the same trade taken much further. The difference is in how distortion is measured. A JAI decoding will not match the original pixel for pixel, but it can show the same things, arranged and lit the same way, from a thousand times fewer bytes.
It works because the decoder brings a prior. A generative model has already learned what harbours, castles and dusk light look like, so the code only has to say which ones, where and how. Speech coding has relied on this for decades: a vocoder sends the parameters of a model of the voice instead of the waveform, and a CELP encoder chooses them by running the decoder inside the encoder and keeping what sounds closest.
The use cases
01 Little bandwidth, enough compute
Spacecraft and deep-space links, ships and aircraft, satellite plans with data caps, rural and disaster-zone networks. Text-only channels such as SMS, packet radio and mesh networks can carry a sentence where they cannot carry a photograph. A receiver with a GPU draws the picture.
02 Local generation
Generating an image is moving onto ordinary GPUs and phones, at well under a second. When decoding is local and free, a browser can pick the JAI source in a <picture> element and never download the raster. Pages get lighter with no change to their markup.
03 An audience of one
Your decoder draws the pixels, so you can ask for them your way: adapted to your culture, softened for content you would rather not see, adjusted for your eyesight, with the text in the picture in your language, or just in a mood you enjoy. The author's description stays the same, and the personalised render is always labelled as such.
04 Longevity and any resolution
A description does not age with its codec and has no fixed resolution. Rendered next year by a better model, or for a larger screen, the same few hundred bytes give a better picture, the way vector graphics outlive the screens they were drawn for.
05 Text all the way down
A JAI image can be searched, translated, compared in version control and edited in a text editor. It carries its own text alternative for screen readers and text-only browsers, and language models and agents read and write it in the medium they already use.
06 Honest by construction
A JAI image is generated by definition. It cannot be mistaken for a photograph of something that happened, and it says so in its format.
Encoding and decoding
Decoding is the generative model turning the description back into pixels. Encoding is the harder half, and it works by analysis by synthesis: describe the image, render the description with the models you want to support, compare each render with the original, revise the words, repeat. The loop stops when the renders match as well as they are going to, or when the time or money set aside runs out.
Like a font stack in CSS, a document can name the models it was tuned for, in order of preference. It carries a core description that every model understands, plus small refinements for particular models, so each one reconstructs the image as well as it can and any other model still gets the core.
jai:A girl holding a red balloon in a park at dusk;prompt/flux=The balloon floats high above her;prompt/imagen=The balloon is glossy;model=flux;model=imagenOne core description, with refinements for two models in the order of preference. Any other model renders the core alone.
The stage we are at
Most people do not have a fast image model on their machine yet, so today's renders usually go to a hosted provider, paid for with an API key. That is good for demonstrations and for personal, audience-of-one renders, and it is why the specification is careful about consent, budgets and what leaves the browser. The format does not depend on it: the same documents decode on a local GPU, and get better as models do.
Where exact pixels matter more than size, such as evidence, medical images, diagrams and screenshots of text, keep the raster. JAI can still travel alongside it as the searchable, editable source.