Real renders, real documents

Every image on this page was rendered once from the JAI document shown beside it and saved, so viewing the page needs no API key. Open "How this was rendered" under any image to see the exact text sent to the model.

A sentence is a complete image

No prefix, no settings. In a context that already says "this is JAI", such as a .jai file or a data-jai attribute, plain prose is a valid document.

The renderer filled in everything the sentence left open: the cobblestones, the leaf shadows, the shutter. That is the trade JAI makes. You decide what matters and say it; the decoder invents the rest.

The same document, in the three forms a program can exchange:

A red bicycle leaning against a whitewashed wall, afternoon sun
{"jai":1,"prompt":"A red bicycle leaning against a whitewashed wall, afternoon sun"}
data:image/jai,A%20red%20bicycle%20leaning%20against%20a%20whitewashed%20wall,%20afternoon%20sun

Same description, different style

Only the style setting changes between these three documents. The renderer, model and seed are the same, which is why the composition holds while the medium changes.

One document, three renderers

The document says it prefers a FLUX model, then SDXL. Preferences are advisory: the reader's viewer chooses among the renderers its user has configured. Here the same document went to three of them.

Note how each renderer read "Dutch golden age" differently. A JAI document fixes the intent, not the pixels. A model value is only ever matched against renderers the reader has set up; it never makes a viewer download or run anything.

Restyle the image already on the page

With mode=transform the page's own image becomes an input. strength says how far the result may move away from it: 0 keeps it, 1 ignores it.

The markup

<picture>
  <source type="image/jai"
          srcset="data:image/jai,The%20same%20harbour%20scene%20as%20a%20pencil%20sketch%20on%20cream%20paper;mode=transform;strength=0.7">
  <img src="harbour.webp" alt="Fishing boats moored in a misty harbour at dawn" crossorigin="anonymous">
</picture>

A viewer may only send pixels the page itself could read: same-origin images, or images served with CORS (hence crossorigin). It never uses the reader's cookies to fetch someone else's picture and send it to a provider.

From a photograph to 465 bytes

An encoder runs the decoder inside the loop: a vision model describes the image, the target models render the description, the renders are scored against the original, and the description is revised. This is NASA's Earthrise, taken by Bill Anders on Apollo 8 (public domain), encoded for FLUX.1 schnell and Stable Diffusion XL with the reference encoder in jai-tools.

The encoded document

The model list is ordered by how well each model reproduced the original. The refinement after prompt/flux= reaches FLUX models only; every other model gets the core description. Download the .jai file (465 bytes, gzip JSON) or the full report with every candidate and score.

The run used Gemini 2.5 Flash to describe, revise and rate, and CLIP embeddings plus a thumbnail colour and layout distance to score. It made 40 renders in about eight minutes for an estimated $0.38. The score blends those measures with the vision model's rating, which is noisy, so the encoder re-rates the current best alongside every new generation. Fidelity is bounded by the decoders: neither model reliably draws a small Earth where it is asked to. The rate, a few hundred bytes, does not change when the decoders get better; the distortion does.

What a reader sees, step by step

A JAI-aware viewer shows the original image (or a tiny placeholder) while a render is on its way, swaps in the render, and keeps the original one click away. Choose a scenario and press Render. This demo plays back saved images; nothing is sent anywhere.

Original

Render Show original Reset

Scenario

generate transform keep no original render fails

Copy