ModelAufgelistetExtern

Pixlvault Anomaly Tagger

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

Quellehuggingface.co/PersonalJeebus/pixlvault-anomaly-tagger/tree/94473f195328ad107d6139b72c3650ee4060f823

ArtOtherBasismodellfacebook/convnext-base-224VersionRevision 94473f1CBewertung

Modellquelle

Art: Other
Dateigröße: 338–341 MB
Basismodell: facebook/convnext-base-224
Version: vRevision 94473f1
Aktualisiert: 10. Okt. 2026
Downloads: 0
Quelle: Hugging Face

Quellenbeschreibung

PixlVault Anomaly Tagger — v238

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

Version: 238 (pixltagger run run-238)
Trained: 2026-10-08
Source commit: df486c6b4e2f
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: 0.708 · micro-F1: 0.662 (across the 14 documented anomaly tags) · overall (all tags) micro-F1: 0.764

Anomaly tag	Precision	Recall	F1	Support
incorrect reflection	0.793	0.768	0.781	95
bad anatomy	0.450	0.502	0.475	215
malformed hand	0.663	0.622	0.642	542
malformed teeth	0.835	0.672	0.745	345
missing nipples	0.947	0.928	0.938	97
malformed nipples	0.661	0.481	0.556	77
waxy skin	0.839	0.656	0.736	462
flux chin	0.838	0.740	0.786	446
malformed foot	0.677	0.548	0.606	199
silicone breasts	0.819	0.686	0.747	86
malformed eyes	0.185	0.848	0.304	171
noise	0.952	0.967	0.959	61
watermark	0.902	0.725	0.804	204
blocky	0.826	0.836	0.831	262

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)

Anomaly tag	Train examples
incorrect reflection	736
bad anatomy	986
malformed hand	6650
malformed teeth	2201
missing nipples	106
malformed nipples	188
waxy skin	3152
flux chin	2466
malformed foot	2581
silicone breasts	714
malformed eyes	1768
noise	976
watermark	4092
blocky	1749

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-238 (commit df486c6b4e2f). Anomaly metrics are on the frozen evaluation set.

Quellen

1 Quelle

Hugging FacePrimärVersion 94473f195328ad107d6139b72c3650ee4060f823

huggingface.co/PersonalJeebus/pixlvault-anomaly-tagger/tree/94473f195328ad107d6139b72c3650ee4060f823

Verifiziert 10. Okt.

Download files

Modellartefakte

2 Artefakte

best.pt

pt · 338 MB · SHA-256 a0f823997050…8c31 · Hugging Face

Download file

pixlstash-anomaly-tagger.safetensors

safetensors · 341 MB · SHA-256 f4e0ed2b8cbc…58c3 · Hugging Face

Download file

Quellenauszüge

3 Auszüge

Readme10. Okt.huggingface.co/PersonalJeebus/pixlvault-anomaly-tagger

--- license: mit base_model: facebook/convnext-base-224 pipeline_tag: image-classification library_name: pytorch tags: - pixlvault - anomaly-detection - multi-label-classification - convnext - pytorch - not-for-all-audiences metrics: - f1 - precision - recall model-index: - name: PixlVault Anomaly Tagger results: - task: type: image-classification name: Multi-label anomaly tagging dataset: type: private name: PixlStash frozen evaluation set metrics: - type: f1 name: Anomaly macro-F1 (documented tags) value: 0.7078 - type: f1 name: Anomaly micro-F1 (documented tags) value: 0.6622 --- # PixlVault Anomaly Tagger — v238 Multi-label image anomaly tagger detecting AI-generation defects (malformed hands/feet, bad anatomy, waxy skin, etc.). Trained with pixltagger. - **Version:** 238 (pixltagger run `run-238`) - **Trained:** 2026-10-08 - **Source commit:** `df486c6b4e2f` - **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: 0.708 · micro-F1: 0.662** (across the 14 documented anomaly tags) · overall (all tags) micro-F1: 0.764 | Anomaly tag | Precision | Recall | F1 | Support | |---|---|---|---|---| | incorrect reflection | 0.793 | 0.768 | 0.781 | 95 | | bad anatomy | 0.450 | 0.502 | 0.475 | 215 | | malformed hand | 0.663 | 0.622 | 0.642 | 542 | | malformed teeth | 0.835 | 0.672 | 0.745 | 345 | | missing nipples | 0.947 | 0.928 | 0.938 | 97 | | malformed nipples | 0.661 | 0.481 | 0.556 | 77 | | waxy skin | 0.839 | 0.656 | 0.736 | 462 | | flux chin | 0.838 | 0.740 | 0.786 | 446 | | malformed foot | 0.677 | 0.548 | 0.606 | 199 | | silicone breasts | 0.819 | 0.686 | 0.747 | 86 | | malformed eyes | 0.185 | 0.848 | 0.304 | 171 | | noise | 0.952 | 0.967 | 0.959...

Readme1. Aug.huggingface.co/PersonalJeebus/pixlvault-anomaly-tagger

--- license: mit base_model: facebook/convnext-base-224 pipeline_tag: image-classification library_name: pytorch tags: - pixlvault - anomaly-detection - multi-label-classification - convnext - pytorch - not-for-all-audiences metrics: - f1 - precision - recall model-index: - name: PixlVault Anomaly Tagger results: - task: type: image-classification name: Multi-label anomaly tagging dataset: type: private name: PixlStash frozen evaluation set metrics: - type: f1 name: Anomaly macro-F1 (documented tags) value: 0.6970 - type: f1 name: Anomaly micro-F1 (documented tags) value: 0.7067 --- # PixlVault Anomaly Tagger — v166 Multi-label image anomaly tagger detecting AI-generation defects (malformed hands/feet, bad anatomy, waxy skin, etc.). Trained with pixltagger. - **Version:** 166 (pixltagger run `run-166`) - **Trained:** 2026-07-18 - **Source commit:** `0525dc55f17e` - **Architecture:** convnext_base, input 768px, ImageNet normalization - **Labels:** 1242 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: 0.697 · micro-F1: 0.707** (across the 14 documented anomaly tags) · overall (all tags) micro-F1: 0.750 | Anomaly tag | Precision | Recall | F1 | Support | |---|---|---|---|---| | incorrect reflection | 0.781 | 0.792 | 0.786 | 72 | | bad anatomy | 0.711 | 0.443 | 0.545 | 183 | | malformed hand | 0.515 | 0.536 | 0.525 | 362 | | malformed teeth | 0.767 | 0.622 | 0.687 | 233 | | missing nipples | 0.957 | 0.746 | 0.838 | 59 | | malformed nipples | 0.804 | 0.587 | 0.679 | 63 | | waxy skin | 0.822 | 0.707 | 0.760 | 406 | | flux chin | 0.803 | 0.741 | 0.771 | 340 | | malformed foot | 0.507 | 0.555 | 0.530 | 128 | | silicone breasts | 0.839 | 0.644 | 0.729 | 73 | | malformed eyes | 0.833 | 0.125 | 0.217 | 40 | | noise | 1.000 | 0.950 | 0.974 |...

Note10. Okt.huggingface.co/PersonalJeebus/pixlvault-anomaly-tagger

Source context: 0 downloads · 1 likes · Pipeline image-classification · Library pytorch · Repo PersonalJeebus/pixlvault-anomaly-tagger