Captcha Lab reads your own challenge images with a CRNN + segmentation ensemble, generates eleven kinds of synthetic test challenges with ground truth, trains better models from your data, and hands real third-party checks to a person through the official widget.
From a single image to a full regression suite: recognise, pick, annotate, evaluate, train and promote — with honest confidence at every step.
A CRNN with CTC decoding reads distorted text end-to-end, while recognition-driven segmentation offers a second opinion. When they disagree, the answer is flagged instead of guessed.
Click, drag, keyboard or touch to select tiles, then compare against model suggestions with undo and redo.
Text, math, colour-text, grids, odd-one-out, counting, click-the-target, slider, rotation, audio and checkbox — each with ground truth and difficulty 1–5.
ZIP import, paste-to-label, review queue, hard-example mining, deduplication and leakage-safe splits. Models are promoted only when they win.
reCAPTCHA, hCaptcha and Turnstile are detected and paused, shown to a person in the official widget, and validated server-side with strict hostname, action, freshness and single-use checks.
Latency percentiles, calibration tables, confusion analysis, model leaderboards, audit logs and Prometheus metrics.
Each stage is timed, bounded by a deadline and validated, so one malformed upload can never take the system down.
Magic bytes, size and decompression-bomb limits, EXIF orientation.
Illumination flattening, white balance, adaptive denoise and sharpening.
Regions of interest, grids, objects and character boundaries.
CRNN, segmentation MLP and CNN tile classifier, with optional canary or shadow models.
Calibrated confidence and disagreement checks. Uncertain answers go to the review queue.
Sign in with your owner account to open the studio.