Prompt Lab

Diagnostics

Checks D1, Google, and xAI credentials without spending anything.

1 · Image set

Load replaces the grid. Add keeps what is already there, so you can build a set from several orders.
Pool not loaded.

2 · Mode

Analysis-only costs about $0.004 per image per arm and takes seconds.

3 · Arms

Arm A

Arm B

If both arms use the same analysis model and effort, the analysis runs once and both arms share it — cheaper, and it removes a confound.

Only the assembly strategies read the per-image selection. Any other strategy ignores it entirely, which is what makes it a clean reference side.

4 · Image models — one row per checked model

5 · Start

Pick images to see the estimate.

Progress

No run in progress.
No run loaded.

Print Prep review

Two versions of Print Prep run in this tab against the image that actually printed, cropped to the 2" square magnet at 640 px. You pick the better side and drag it until it is right; the gap between the recommendation and your correction is the measurement.

UI — · if this is not the build you just deployed, the browser is serving a cached file and any session you run measures the old one.

Square magnets only — rectangles and mosaics print at other sizes and are excluded.

No images picked yet.

Leave both on Production for a diagnosis run — that measures where production stands before anything is changed.

Residuals

Not loaded.

Sharpen calibration

Production reads Laplacian variance off a 1000 px analysis canvas. The magnet prints at 640 px, and scaling a crop to that size changes how sharp it reads. This asks which of the two measures better predicts the level you actually chose.

Not loaded.

Pet eyes — catchlight detection

Two Moondream 3.1 calls per image on Workers AI — point for eyes, detect for faces — at roughly a twentieth of a cent for the pair against $0.10 for an AI enhancement run. Cloudflare's published median is ~145ms and ~160ms; this session measures what you actually get. The model gives a neighbourhood, then local code snaps to the dark disc and measures the iris, so precision does not depend on the model being pixel-accurate.

Overlay: amber ring = where Moondream pointed · green ring = the iris the pixels confirmed · blue box = Moondream's face · grey dashed box = TinyFaceDetector's face. A point with nothing dark under it is dropped rather than lifted.

The face comparison is the second question. TinyFaceDetector is a human-face model, which is why pet photos fall through to the frame histogram and why candidate 9c had to exist. If Moondream finds pet faces reliably, that changes which branch of print prep a pet takes.

No images picked yet.

Runs

WhenRunImagesCellsResult