Cerberus Image v3

Run it

pip install -r requirements.txt
python3 service.py
curl -X POST http://127.0.0.1:5972/check \
  --data-binary @photo.jpg

{"nsfw": false, "score": 0.12, "labels": [],
 "nude_score": 0.05, "open_score": 0.12,
 "model": "cerberus-image v3"}

Spec

• SignalsPart detector + broad classifierAn image is blocked if either trips: explicit-part detection or whole-image NSFW score.
• Cutoffsnude ≥ 0.35 · open ≥ 0.10Grid-tuned on 12,144 reviewer votes. Per-request override via X-Nude-Threshold / X-Open-Threshold.
• Accuracy0.86 global · 0.91 porn · 0.81 furryAgreement with labels on the tuning set, see Training data below. Drawn content is the hard slice.
• Runs onAny CPU boxPython + onnxruntime. No GPU, no API keys, nothing leaves the machine.

Training data

Cutoffs are grid-searched against 12,144 labeled points: reviewer Correct/Wrong votes on live verdicts, plus rating-derived auto-labels for scale. The vote store doubles as the regression set, so retrains re-check the corpus under new cutoffs.

• Most labels are weak auto-labels. The human-reviewed subset agrees with the model about 72 percent, so read accuracy as agreement with imperfect labels, not ground truth.
• Drawn content is the hard slice for the part detector (0.81 vs 0.91). The broad classifier carries that slice, which is why the OR verdict exists.
• Thresholds drift as the content mix changes. Cutoffs are re-tuned on a schedule, not once. No fairness or bias evaluation has been done.