YOLO-S (Fashion Item Detector): 46 garment and trim categories, and where to run it

Start with the licence, because it is different from almost everything else here. This fine-tune carries CC BY 4.0 rather than Apache 2.0, which means commercial use is fine but attribution is required. If the output feeds a public product catalogue, that is a credit line someone has to write. Self-hosting is otherwise easy: PyTorch, the Hugging Face transformers library, and a small model. In your browser it runs on CNAPS Studio, included in the basic plan.

What it does that no general detector can is name the parts of a garment. A general model sees a person and a shirt. This one sees a shirt, and also its collar, its lapel, its sleeve, its pocket, its buttons and its zip, across 46 fashion categories covering both what you are wearing and how it was made.

The cost is that it knows nothing else. Feed it a street scene and it will look for clothing, because clothing is all it has labels for. It also publishes no accuracy score of any kind. One fashion detector sits in the CNAPS Studio catalog, and this is it.

What it actually is

The same small vision transformer as the general YOLO-S, retrained for a different label set. Twelve layers, twelve attention heads, 384 hidden dimensions, 16 by 16 pixel patches, and 100 object queries. The only structural change is the classification head, which now chooses among 46 fashion classes rather than 80 everyday ones.

The training data is the interesting part and explains the licence. It was fine-tuned on FashionPedia, a public dataset of 46,800 images annotated with those 46 categories, and the CC BY 4.0 terms come from that dataset rather than from the model code.

The 46 categories split into two useful groups. Garments and accessories cover the obvious ground: shirt, t-shirt, sweater, cardigan, jacket, vest, coat, cape, jumpsuit, pants, shorts, skirt, plus glasses, hat, headband, tie, glove, watch, belt, tights, sock, shoe, bag, scarf and umbrella. Construction details are the unusual half: collar, lapel, epaulette, sleeve, pocket, neckline, buckle, zipper, applique, bead, bow, flower, fringe, ribbon, rivet, ruffle, sequin and tassel.

ArchitectureSmall vision transformer, no convolutional backboneEncoder12 layers, 12 attention heads, 384 hidden dimensionsObject queries100, which is also the maximum items per imageClasses46 fashion categories, garments and construction detailsBase modelYOLO-S, pre-trained on ImageNet-1kFine-tuning dataFashionPedia, 46,800 annotated imagesOutputLabel, box as left/top/right/bottom, confidence 0 to 1LicenceCC BY 4.0, attribution requiredPublished accuracy scoreNone

What is published, and what is not

There is no accuracy score. The general YOLO-S it was built from publishes 36.1 average precision on COCO, but no equivalent figure exists for this fine-tune on fashion data. No evaluation set, no per-category breakdown, nothing to check.

Our own documentation offers confidence bands by category instead: strong on large garments, good on accessories, weaker on small details like sequins and beads. Those are our descriptions rather than measurements, and we wrote them, so weigh them as such.

The practical consequence is the same as for any unmeasured model, and it is sharper here because the label set is fine-grained. A model that must tell a lapel from a collar and a rivet from a button has many ways to be subtly wrong. Before this tags a live catalogue, run a few hundred of your own product photos and count the mislabels yourself, particularly on the construction-detail classes.

How it compares to the other detectors here

YOLO-S (Fashion)YOLO-SRF-DETR MediumLooks for46 fashion categories, garments and trim80 everyday categories80 everyday categoriesKnows a collar from a lapelYesNoNoSizeSame as YOLO-S, 30.7M30.7M33.7MPublished scoreNone36.1 COCO average precision54.7 COCO average precisionLicenceCC BY 4.0, attribution requiredApache 2.0Apache 2.0

The score row cannot be compared across these columns even where numbers exist, because a fashion label set and a general one are different tasks. A model scoring well on COCO tells you nothing about whether it can find a zip.

Choosing is about the label you need. Reach for this one when the answer must name a garment or a garment part, which is most e-commerce cataloguing and inventory work. Reach for RF-DETR Medium or YOLO-S when you only need to know that a person or a bag is present, since both are measured and both are more accurate at those broad categories. And when the thing you want has no label anywhere, SAM 3.1 (Scene) takes a text phrase such as the red stitching on the cuff and returns a mask.

What to chain it with

One image in, labelled boxes with confidence scores out. The labels being fashion-specific is what makes the chains different from a general detector's.

Two are worth building. For cataloguing, Image Crop by Class cuts each detected garment out as its own image and BLIP (Image Description) writes a description of it, which turns one model shot into a set of per-item listings. For editing, Image Masker by Class turns a garment box into a mask and QWEN-Inpaint replaces just that garment, which is how you show one jacket in four colourways without a reshoot.

Open YOLO-S (Fashion Item Detector) in CNAPS Studio, run one of your own product photos, and check the construction-detail labels specifically, since those are where an unmeasured model is most likely to be confidently wrong.

Sources

• huggingface.co/valentinafeve/yolos-fashionpedia and the author's fine-tuning repository for the base model, the fine-tuning approach, and the absence of any published accuracy figure.
• The FashionPedia dataset for the CC BY 4.0 licence and the 46 categories.
• huggingface.co/hustvl/yolos-small for the shared architecture and the general model's 36.1 average precision.
• docs.cnaps.ai/ai-model-hub for the CNAPS Studio input and output, the full category list, and the qualitative confidence notes.