pixlvault-anomaly-tagger

PixlVault Anomaly Tagger — v255

Multi-label image anomaly tagger detecting AI-generation defects (malformed hands/feet, bad anatomy, waxy skin, etc.). Trained with pixltagger.

• Version: 255 (pixltagger run run-255x)
• Trained: 2026-10-22
• Source commit: 9014a0bc16a4
• Architecture: convnext_base, input 768px, ImageNet normalization
• Labels: 1876 tags (14 documented anomaly tags + descriptive tags)

Intended use & limitations

Built to flag defects in AI-generated images for automated QA / down-ranking in PixlVault. It is not a moderation tool for real photographs and was not validated on natural images — training is synthetic-heavy, so expect a domain gap. Outputs are per-tag probabilities thresholded for precision; rare tags carry wide run-to-run variance on the small evaluation set. Adult content: the model detects anatomical anomalies including nudity-related defects, hence the not-for-all-audiences tag.

Anomaly performance (frozen evaluation set)

Anomaly macro-F1: 0.714 · micro-F1: 0.671 (across the 14 documented anomaly tags) · overall (all tags) micro-F1: 0.768

Per-class production thresholds are stored in pixlstash-anomaly-tagger_meta.json under label_thresholds (precision-biased, min precision 0.75). Apply sigmoid(logits)[tag] >= label_thresholds[tag] per tag.

Documented anomaly tags

incorrect reflection, bad anatomy, malformed hand, malformed teeth, missing nipples, malformed nipples, waxy skin, flux chin, malformed foot, silicone breasts, malformed eyes, noise, watermark, blocky.

Training data (this run, anomaly tags)

Usage

import json, torch
from safetensors.torch import load_file
from torchvision.models import convnext_base

labels = json.load(open("labels.json"))
meta = json.load(open("pixlstash-anomaly-tagger_meta.json"))
thresholds = meta["label_thresholds"]
model = convnext_base()
model.classifier[2] = torch.nn.Linear(model.classifier[2].in_features, len(labels))
model.load_state_dict(load_file("pixlstash-anomaly-tagger.safetensors"))
model.eval()
# probs = torch.sigmoid(model(x)); tag fires when probs[i] >= thresholds[labels[i]]

---

Model card generated by pixltagger pixl promote from run run-255x (commit 9014a0bc16a4). Anomaly metrics are on the frozen evaluation set.