Incant

A wizard’s canvas for the Daylight DC-1. Draw a sketch with the stylus, hold the wand and speak a spell, release — the sketch transfigures into a finished image.

Draw. Speak. Transfigure.

Ornate cabinet is on by default — carved oak, brass, candles, the parchment in a sunken well. Grimoire → Ornate cabinet turns it off for the flat folio.

A quill sketch of a cottage, then the same composition after the spell:

A mountain trail:

ChatGPT Sketch (@Sketch) is a ChatGPT product surface, not a separate developer API. There is no dedicated “sketch mode” endpoint.

The same capability is available to apps through the Images API:

• Model: gpt-image-2.5-flare (fast) or gpt-image-2.5-sunburst (precise)
• Endpoint: POST /v1/images/edits
• You send the sketch as the reference image and the spoken (or typed) spell as the prompt

That is the official “turn a drawing into a realistic image” workflow. Incant uses it. An iframe of ChatGPT is unnecessary.

Voice uses Gemini 3.5 Live Transcribe (gemini-3.5-transcribe-live) over the Live API, push-to-talk / manual VAD — hold the wand, speak, release to finalize.

The live app is a portrait, grayscale-first parchment UI with software palm rejection (canvas ignores finger/touch; only pen and mouse draw). The DC-1’s Wacom EMR layer already rejects palms in hardware; Incant double-filters.

1. Open the web app in the DC-1 browser, or install it as a PWA.
2. Open Grimoire and paste an OpenAI key and a Gemini key. Keys stay on the device.
3. Draw. Hold the wand. Speak. Release.

There is also a native Jetpack Compose app in android/ for sideloading:

• Custom SketchView with stylus pressure and palm rejection (TOOL_TYPE_FINGER ignored)
• Hold-to-talk wand
• Same OpenAI + Gemini wiring

Open android/ in Android Studio, let Gradle sync, run on the DC-1.

• Ornate cabinet (on by default) is the cereal-box relic: oak, brass, depth. Off is the flat folio, better if you want a quieter page.
• One visage first. Casting generates one image so it returns quickly. Another weaves a new interpretation of the same sketch + spell. Previous results collect as echoes along the bottom.
• Fourfold visages (Grimoire toggle, off by default) fires four in parallel. The first to return becomes the large image; the rest fill in as echoes. Four tiny thumbnails of similar images are hard to tell apart on Live Paper, so the large image stays primary.
• Compose for Live Paper (on by default) asks the model for strong value contrast so the result still reads in grayscale.
• You can always type the spell if the microphone is unavailable.

Keys never leave the device except to those APIs.