Updraft

Browser-based AI image tools that run entirely on your machine. Two features today: a super-resolution Upscaler and Background Removal. Models run client-side via ONNX Runtime Web; images never leave the page.

Run it

npm start        # serves the static site on :8081

No build step. It's a static site of ES modules and Web Components, wired through an import map in index.html. Open it, pick a feature, drop an image.

How it works

Upscaler. Drop an image, optionally crop, select a model, hit Upscale, click to compare, download.

The work is a tiled pipeline (features/upscaler/upscale-pipeline.js):

• The image is cut into overlapping tiles so big images fit in VRAM. Each tile is run through the model; tiles are feathered back together to hide seams (engine/tiling.js, engine/upscaler-engine.js).
• The pipeline is a list of optional steps over a shared context: base upscale, a second model for side-by-side comparison, an all-image blend pass, and face detect → per-face enhance. Each step declares shouldRun(ctx), so disabled passes cost nothing.
• Models come from a registry of built-in ONNX files (model-registry.js), plus your own: upload an .onnx, it's inspected for I/O shape, scale, and precision (inspect-onnx.js), then cached in the browser Cache API and treated like any built-in.

Background removal is simpler: one segmentation model produces an alpha mask that's composited over transparency (features/bg-removal/).

Backends

Engines ask lib/backend.js for a session with an intent — gpu or cpu — and get back a session plus a label of what actually ran. The web path uses ORT-Web's WebGPU EP, falling back to WASM. Fallbacks and the realized backend surface in the status bar (lib/backend-events.js).

Desktop (optional, faster)

An Electron wrapper in desktop/ routes inference through onnxruntime-node with native execution providers — CoreML on macOS, CUDA on NVIDIA, DirectML on Windows — for a 3–10× speedup. The renderer is unchanged: lib/backend.js dispatches to native only when Electron's preload exposes window.__nativeOrt. See desktop/README.md.

Layout

index.html              # entry: import map + feature switcher
features/upscaler/       # pipeline, tiling/inference engines, UI
features/bg-removal/     # segmentation engine + UI
components/              # shared Web Components (cropper, sliders, status)
lib/                     # backend dispatch, model cache, ORT loader, canvas utils
desktop/                 # Electron native-acceleration wrapper
vendor/                  # ORT-Web, PicoCSS, Font Awesome (no package manager)