Stylor

Release

Introducing Stylor Segment v1

Our first segmentation model, built for fashion. It finds every garment, shoe, bag and piece of jewellery in a photo and returns a precise mask for each one.

Read the docsGet an API key
A woman in a studio in a beige trench coat over a striped t shirt, blue jeans and brown loafers, with gold hoop earrings, sunglasses and a tan handbag
A man in a studio in a navy blazer over a white t shirt, beige trousers and white sneakers, with a watch and sunglasses
A woman in a studio in a red sleeveless midi dress and black pumps, with a gold necklace, earrings, a bracelet and a black clutch
A man in a studio in a cream cable-knit sweater, navy shorts, white socks and sneakers, with a navy cap and a watch
A woman in a studio in a black leather jacket over a white t shirt, wide camel trousers and white sneakers, with gold necklaces, earrings and a black shoulder bag
A woman in a studio in a denim jacket over a white sundress and tan sandals, with a straw hat, sunglasses and a woven tote bag

Try Stylor Segment v1

Each photo below is shown with the masks the model returned. Point at a garment to lift it out, or try a photo of your own:

A man walking on a city street in a long camel coat over a grey hoodie, black jeans, white sneakers and a navy beanie, with a canvas tote bag

6 items0.19 s

Jewellery and small pieces

Jewellery is small and easy to miss. With zoom on, Segment v1 takes a closer second look at the photo and returns each necklace, ring and earring as its own item:

A woman with grey hair tied back in a cream sleeveless knit top, with pearl drop earrings, a pearl necklace, a pearl bracelet and a gold ring
  • Necklace
  • Bracelet
  • Earring
  • Ring

Evaluating Stylor Segment v1

We tested Segment v1 against six other open models on four sets of fashion photos, 799 photos in all. On every photo we compared each mask a model returned with one a person drew. The score, item IoU, is how closely the two match, from 0 to 1.

Segment v1 scores highest on all four test sets:

Mask qualityItem IoU, mean of the four test sets (n=799)
  1. Stylor Segment v10.795
  2. Stylor Segment v1, with zoom0.799
  3. SAM 30.733
  4. SegFormer-B3 (Fashionpedia)0.643
  5. Grounded-SAM 20.608
  6. SegFormer-B2 (clothes)0.570
  7. FASHN human parser0.521
  8. YOLOE-11L0.401

SAM 3 comes closest throughout. The lead is not the same everywhere: it is widest on our own benchmark and on Fashionpedia, and narrowest on DeepFashion2.

Speed

Segment v1 is also fast. It takes 0.15 s for a photo, or 0.32 s with zoom. SAM 3, the only model close to it on quality, takes 2.45 s:

Quality against timeAverage item IoU against the time for one photo. Every model ran on the same hardware.
0.30.40.50.60.70.80.930 ms100 ms300 ms1 s3 sTIME FOR ONE PHOTOSegment v1with zoomSAM 3SegFormer-B3Grounded-SAM 2SegFormer-B2FASHNYOLOE

Four models are faster than Segment v1. Each of them scores at least 0.15 lower.

Side by side with SAM 3

Four photos where the two models disagree, as each answered them:

SAM 35 items · 2.6 s
SAM 3’s masks on a man walking on a city street in a long camel coat over a grey hoodie, black jeans, white sneakers and a navy beanie, with a canvas tote bag
Stylor Segment v16 items · 0.2 s
Stylor Segment v1’s masks on a man walking on a city street in a long camel coat over a grey hoodie, black jeans, white sneakers and a navy beanie, with a canvas tote bag

Results in full

Item IoU for every model, and the time each took for one photo. The best figure in each column is in bold:

ModelAverageTime
Stylor Segment v10.7950.15 s
Stylor Segment v1, with zoom0.7990.32 s
SAM 30.7332.45 s
SegFormer-B3 (Fashionpedia)0.64378 ms
Grounded-SAM 20.6080.29 s
SegFormer-B2 (clothes)0.57069 ms
FASHN human parser0.5210.13 s
YOLOE-11L0.40138 ms

Methodology

  1. 1The score. Each hand-drawn item is matched to one of the model’s masks and scored by how much the two overlap. An item the model missed scores 0. A photo’s score is the mean over its items, and a test set’s is the mean over its photos.
  2. 2The same prompt for every open-vocabulary model. SAM 3, Grounded-SAM 2 and YOLOE were each given the same 38 garment names. Grounded-SAM 2 and YOLOE were run at the detection threshold that scored best for them on a sample of our benchmark.
  3. 3DeepFashion2’s outlines are loose. They take in hands, phones and hair, so every model scores lower there than its masks deserve.
  4. 4Time. Every model ran on the same hardware, one photo at a time, timed from the decoded photo to the finished masks.
  5. 5The test sets. Our benchmark is not public. The other three are: 100 of Fashionpedia’s validation photos, picked for one person in a layered outfit, and 300 each from Clothing Co-Parsing and DeepFashion2. A pair of shoes, socks or gloves counts as one item throughout.

Availability

Stylor Segment v1 is available today on the Stylor API. If you’re a developer, use stylor-segment-v1 on the segment endpoint. The docs have the full reference, every label, and code for using a mask in JavaScript and Python.

Read the docsGet an API key