jev-paint

A small local app that turns Jev's pixel probability distributions into paintings.

Run

Requires Python 3.9+ and a modern browser with module workers and OffscreenCanvas (current Chrome, Edge, Firefox, or Safari). No packages, build step, or Node installation needed.

python3 server.py

Open http://127.0.0.1:8791, open Settings, and paste your Jev API key. Choose a representation and grid size, then enter a prompt. Each painting joins an in-memory carousel. The key is saved in this browser’s local storage; clear the key field to forget it. Paintings are cleared when you reload. Generation uses your TypeSafe account. Stop the helper with Ctrl+C. Use python3 server.py --port 8795 if the default port is occupied.

All question construction, batching, probability processing, and painting run in the browser. A Python standard-library helper serves the files and forwards requests to the fixed Jev endpoint because its API currently disallows browser CORS origins. It binds only to loopback, checks the request origin, and neither stores nor logs keys or prompts. No environment variables are required.

Rendering

The same algorithm handles palette, HSL, binary RGB, and silhouettes. HSL/RGB channel probabilities are combined assuming independence; Jev does not supply a joint distribution. Spatially correlated categorical sampling supplies stroke pigments, mean color supplies the underpainting, and entropy controls relief. Strokes follow local luminance contours; a height map supplies impasto lighting. This is an artistic transformation, not a lossless probability chart.

Painting runs in a module worker at 560×560. Pigments are sampled lazily and cached, and stroke segment geometry is precomputed. The completed textured painting fades in over 550 ms. Network/model time is separate from rendering time.

Code

• server.py: static files and fixed-destination API forwarding
• web/jev.mjs: typed questions, four concurrent batches, validated results
• web/art.mjs: probability representations
• web/renderer.mjs: deterministic painting algorithm
• web/paint-worker.mjs: off-main-thread rendering
• web/app.mjs, index.html, style.css: gallery and prompt UI

Checks (Node 18+)

node --test tests/jev.test.mjs
node tests/benchmark.mjs
python3 -B -m unittest discover -s tests -p 'test_*.py'

Four small recorded probability fixtures are retained for offline rendering checks. These contain public art prompts and model probabilities, with no credentials or account metadata. Tests and benchmarks run offline and incur no API charges.

Privacy and local use

Your API key is stored in browser localStorage for this origin, not encrypted. Clear the key field to remove it. Use your own computer/browser profile. The local helper holds the key only while forwarding a request; prompts and keys are sent only to TypeSafe. Changing the port or using localhost instead of 127.0.0.1 uses a separate browser storage origin.

This is a local creative experiment, not a publicly hosted multi-user service. Keep the helper bound to loopback. Model requests cost tokens; larger grids make more requests. Rendering time and model/network latency are independent.

Contributing

Small fixes and renderer experiments are welcome. Run the checks above before opening a pull request. Preserve the offline fixtures, avoid committing API keys or personal generated outputs, and describe visual changes with before/after images when useful. Browser code uses ES modules with no runtime dependencies.

License

MIT — see LICENSE. Jev is a TypeSafe service and requires your own API access; this project is independent of TypeSafe.