CaptchaKrakenV2-CNN-lite

The lite CaptchaKraken v2 element detector: the CPU-friendly sibling of CaptchaKraken/CaptchaKrakenV2-CNN-large. It finds a captcha board's interactive elements (grid, cell, item, draggable, drop_target, slider, text_input, button, click_area, fill, checkbox).

The CaptchaKraken V2 CNN family

The v2 element detectors: they find a board's interactive elements and draw the numbered marks the v2 pointer models (CaptchaKraken/SunlightV2-0.8B, CaptchaKraken/TwilightV2-2B, CaptchaKraken/MidnightV2-4B) read.

Repo	What
CaptchaKraken/CaptchaKrakenV2-CNN-large	detector.onnx (the client's standard CPU detector) and the Mask R-CNN ResNet-18 training checkpoint detector_r18.pt
CaptchaKraken/CaptchaKrakenV2-CNN-lite	detector-lite.onnx, the MobileNetV3 lite CPU detector for one or two vCPUs — this repo

detector-lite.onnx: the lite CPU detector

The same 11 classes on a MobileNetV3-Large FPN (128 channels, GroupNorm, 1024-wide box head), for one or two vCPUs: 80-107 ms a board on one core, against 520-606 ms for detector.onnx. Same ONNX interface; model metadata arch is lite. The client picks it with detector: "lite" in captchakraken.config.json, or under auto on a machine with two usable CPUs or fewer.

• Trained 2026-11-01 from ImageNet weights: 17,274 steps on the 46,065 generated boards (149 generators + checkbox widgets) that detector.onnx in CaptchaKrakenV2-CNN-large was trained on (short side 384-512), then 4,000 steps on the same boards with the checkbox widgets drawn one vendor at a time. Per-class numbers and both runs' settings in detector-lite.onnx.metrics.json.
• Val (4,606 boards): mask mAP@0.5 0.926, box mAP@0.5 0.930; checkbox mask AP50 1.000, draggable 1.000, grid 1.000, cell AP50 0.970.
• Real captures: every reCAPTCHA, hCaptcha, Prosopo and GeeTest grid board read as detector.onnx reads it; 132/132 held-out widget tick boxes found and 42/42 ticked widgets left alone.
• Limit: a box that fills the whole picture is not found. GeeTest v4's "click to verify" bar is that case; the client clicks it through its DOM selector and never asks.

The interface (opset 17; input image, float (3, H, W) in 0-1; outputs boxes, labels, scores, masks; class names in the model metadata classes) is the same as detector.onnx in CaptchaKrakenV2-CNN-large.

Results across the family

These detectors do not solve captchas themselves: they draw the numbered marks the v2 pointer models read. The table compares those pointer models, so you can see what each size solves.

Every v2 pointer size, on the same 1,067 hand-labelled real boards, served through vLLM 0.25.1 on an RTX 5080. Percent of boards solved; higher is better. q4 is 4-bit GPTQ (W4A16); the others are bf16.

Plain (the bare board):

	0.8B	0.8B q4	2B	4B	4B q4
All (1,067)	53.5%	53.9%	72.6%	75.4%	75.1%
Click (674)	60.4	61.6	76.3	78.3	77.9
Grid (223)	31.8	29.6	75.8	78.9	78.9
Video (158)	56.3	57.0	54.4	59.5	59.5
Type (12)	33.3	33.3	50.0	50.0	50.0

Marked (the CNN detector's numbered marks drawn on the board), stills only, on the boards all five runs marked:

	0.8B	0.8B q4	2B	4B	4B q4
All (904)	50.2%	49.8%	70.5%	71.6%	71.0%
Click (673)	58.5	58.8	73.3	73.7	73.6
Grid (223)	26.0	22.9	62.3	65.5	63.7
Type (8)	25.0	37.5	62.5	62.5	62.5

• 2B solves about 19 points more than 0.8B on the plain boards, split about evenly between clicks and grids. Grids are the sharpest jump (31.8 to 75.8). 4B adds about 3 more.
• Video barely follows size: 0.8B is ahead of 2B there, and 4B leads by about 3 points.
• 4-bit moves the totals by less than a point. Its largest cost is on marked grids (2 to 3 points). A 2B q4 has not been measured or published.

How far to trust the gaps:

• The 0.8B was trained on an earlier build of the data (2026-10-30), so not all of its gap is model size.
• Compare sizes on the plain table. Marked scores also depend on the detector, and the 4B card names a different detector checkpoint from the 0.8B and 2B cards.
• The 4B run did not score marked video, so marked video is left out for every size. That is why these marked totals differ from the ones quoted elsewhere on the cards, which cover different board sets.
• Type has only 12 boards plain and 8 marked, so treat those rows as rough.
• MidnightV2-4B's own gate results were run through transformers, so use this table, not those, to compare it with the other sizes.

Per-board results: 0.8B, 0.8B q4, 2B, 4B, 4B q4.

License: LGPL-3.0.