Pasma

Clothing recognition that stays on your device.

20 garment types · 6 categories · 21 visual material scores

By Vestes Archive · Apache-2.0 · Python / ONNX / CoreML

Pasma is a family of compact vision models for recognizing a single clothing item in an everyday photograph. It helps wardrobe and fashion applications suggest the garment type, category and likely visual materials, without a network request or a separate background-removal model.

Built by Vestes Archive, the digital atelier behind this project, Pasma brings a small part of our clothing intelligence to the open-source community. Its name refers to a counted bundle of threads: individual visual clues brought together into a structured result.

This is the v1.0 candidate release. Weights and a runnable inference example are included. Suggestions should remain user-correctable; phone-specific validation and some export-parity work are still pending.

Try Pasma in your browser

Open the interactive demo →

Choose Small, Nano or Pico and try a photo of one garment. The demo runs ONNX locally in your browser: your photo is not uploaded. The first run downloads the model and runtime.

At a glance

• Small enough to bundle: the Small and Nano FP16 ONNX graphs are about 10.32 MiB each.
• Made for wardrobe photos: 90.75% type Top-1 on the 854-image test_real slice; 89.12% on all 1,103 held-out real images for Small-320.
• Useful alternatives: Small-320 reaches 97.10% Top-3 on the combined real set, allowing an app to ask the user to choose when the first suggestion is uncertain.
• Local inference: after downloading the model, no API key, image upload or per-image API fee is needed.
• Permissive release: project code and fitted weights are distributed under Apache-2.0 with retained upstream MIT/Apache notices.

Quick start

Python 3.11+ is recommended. No PyTorch installation is needed for ONNX inference.

git clone https://github.com/prikol5ist-art/Pasma.git
cd Pasma
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python examples/classify.py /path/to/garment.jpg

The example uses Small-320 FP32 on CPU and returns the top three types, a category consistent with the selected type, and thresholded visual-material suggestions. To use a smaller-resolution graph:

python examples/classify.py garment.jpg --model models/pasma_nano_256_fp32.onnx

Download the published artifacts directly from the Hub:

pip install huggingface_hub
hf download Prikol5ist/pasma --include "models/*" "examples/*" "requirements.txt" --local-dir pasma
cd pasma
pip install -r requirements.txt
python examples/classify.py garment.jpg

Source and issue tracker: GitHub.

Minimal inference, with already prepared RGB float32 NCHW input:

import onnxruntime as ort

session = ort.InferenceSession(
    "models/pasma_small_320_fp32.onnx",
    providers=["CPUExecutionProvider"],
)
coarse, types, materials = session.run(
    ["prob_coarse", "prob_type", "prob_material"], {"image": image_tensor}
)

The complete example handles EXIF orientation, aspect-ratio-preserving bilinear letterbox, gray padding and conversion to [0,1]. ImageNet normalization is inside the exported graph. Do not apply it again. See deployment notes for the exact contract.

Choose your model

All three variants expose the same labels and output schema. MiB means 1,048,576 bytes. Sizes below are measured ONNX files, rather than estimates from parameter count.

Variant	Input	Parameters	FP16 / FP32	Combined Top-1	Combined Top-3
Pasma-Small-320	320 × 320	5.296M	10.32 / 20.37 MiB	89.12%	97.10%
Pasma-Nano-256	256 × 256	5.296M	10.32 / 20.36 MiB	87.94%	97.01%
Pasma-Pico-256	256 × 256	8.557M	16.40 / 32.69 MiB	87.49%	96.83%

Start with Small-320 for recognition quality. Nano lowers input resolution using the same backbone family. Pico uses ConvNeXt and was the fastest native CoreML candidate on the measured M2 host; actual device/runtime results may differ.

Browse weights and metadata · Model card · Training source and lineage

The bundle includes FP32/FP16 ONNX, PyTorch safetensors and zipped FP16 CoreML packages. Pico also includes an experimental 8.71 MiB conv-only INT8 graph with significant probability drift; it is not the default. Check downloads with:

(cd models && shasum -a 256 -c checksums.sha256)

How it compares

These are recorded configurations evaluated against the same Pasma type taxonomy. test_real contains 854 held-out source-derived real photos; bench_test contains 249 curated home/catalogue images. Combined is their concatenation, not an independent third test. Scores are type Top-1 unless stated otherwise.

Model / configuration	test_real	bench_test	Combined	Combined macro F1
Pasma-Small-320, fine-tuned + distilled	90.75%	83.53%	89.12%	87.01%
Pasma-Nano-256, fine-tuned + distilled	90.05%	80.72%	87.94%	85.44%
Pasma-Pico-256, fine-tuned + distilled	89.11%	81.93%	87.49%	84.84%
SigLIP-Base-Patch16-256, zero-shot	78.22%	87.15%	80.24%	76.95%
FashionCLIP ViT-B/32, zero-shot	75.53%	82.33%	77.07%	73.96%
MobileCLIP-S1, zero-shot	76.93%	80.72%	77.79%	75.27%
OpenCLIP ViT-B/32, zero-shot	65.69%	64.66%	65.46%	64.31%
SigLIP2-SO400M, fine-tuned teacher	94.38%	92.37%	93.93%	92.70%

Small-320 leads the strongest tested zero-shot baseline by 8.88 percentage points on the combined set. SigLIP-Base wins on bench_test, so the ranking depends on the image slice. These comparisons measure task-specific training against zero-shot inference, not an equal-training ablation. The teacher is a scale reference, not a mobile deployment option. Third-party comparison weights are not redistributed.

Small-320's combined Top-1 95% interval is 87.4–90.9%, from 2,000 image-level bootstrap resamples. This interval does not capture correlated items, label uncertainty or training-seed variation. Headline Pasma scores are local PyTorch FP32, without TTA or the exploratory BLOUSE bias; exported precision variants have separate parity results.

Local versus cloud

On bench_test, the recorded Qwen3-VL-30B-A3B endpoint scored 91.97% Top-1, compared with Small-320's 83.53%. Its measured p50 request time was 1,796 ms, including network and provider processing. Small-320's native CoreML p50 was 5.85 ms on an M2 MacBook Air, excluding image preparation and startup.

The cloud model saw 1024-pixel JPEGs; Small used a 320-pixel letterbox. This illustrates an accuracy/operations trade-off, not equivalent input or timing conditions. Pasma is useful when offline execution, local photos and avoiding recurring API calls matter.

Measured Apple performance

Model	CoreML CPU + Neural Engine p50 / p95
Small-320	5.85 / 6.49 ms
Nano-256	4.21 / 4.97 ms
Pico-256	2.27 / 2.96 ms

These are Mac M2 measurements, not iPhone or Android benchmarks. Batch one, 5 warm-ups, 30 timed model calls; decoding, preprocessing and model loading are excluded. CPU fallback is allowed. Android/Web execution, phone latency and peak activation RAM still require validation. Some FP16/CoreML/INT8 configurations exceed the original probability-parity tolerance; see format-specific caveats.

The full benchmark report, detailed evaluation card, technical report and machine-readable evidence accompany this release. They include all saved benchmark sections, per-class results, confusion matrices, confidence intervals, calibration and export-parity diagnostics. Private evaluation images and raw cloud responses are not included.

What the outputs mean

Output	Vocabulary	Interpretation
Garment type	20 types	T-shirt, shirt, blouse, sweater, hoodie, pants, jeans, shorts, skirt, jacket, coat, vest, dress, shoes, boots, sneakers, sandals, accessory, hat, scarf
Category	6 categories	Top, bottom, outerwear, footwear, full body, accessory
Visual material	21 independent scores	Appearance suggestions such as denim, fleece, leather-like surfaces and other material families

Use labels_order.json as the index authority. Category probabilities are sums of type probabilities; derive the displayed category from the selected type to keep the pair consistent. Material scores are independent and do not sum to one.

Visual appearance is not fibre composition. Material mAP for Small is 0.664 on the curated 249-image benchmark, averaging the 17 of 21 classes with positive support. It cannot certify cotton content, fibre percentages, authenticity or care instructions.

Known limits

• BLOUSE versus SHIRT is difficult: Small recalls 11 of 23 combined held-out blouses (47.8%); 21 of 23 are in Top-3. Keep alternatives and manual correction available.
• Closed-set predictions have no explicit unknown class. Multiple garments, unusual crops and unfamiliar items can produce confident mistakes.
• Type labels are source/filename-derived; material targets include curated expected sets. No complete independent human-adjudication record or demographic/device fairness study is supplied.
• Source images, private manifests and teacher checkpoints are not included. Full retraining/evaluation cannot be replicated from a bare checkout alone. The historical headline evaluation used EXIF correction off; the deployment example corrects EXIF.
• Pasma v3 predicts type/category/material appearance. It does not provide segmentation, weather, style, dating or 3D fit predictions.

About Vestes Archive

Vestes Archive is our digital atelier for clothing, personal archives and machine-assisted curation. We build tools that help people understand and care for what they already own. Pasma is the open-source clothing-perception model within that work.

Website: vestesarchive.dev · Contact: team@vestesarchive.dev

License and contributions

Project code and fitted weights are released under the Apache License 2.0, retaining upstream notices. Small/Nano derive from MIT-licensed EdgeNeXt weights; Pico derives from Apache-2.0 ConvNeXt weights. See third-party notices and lineage. The project license does not grant rights to source datasets or third-party baseline weights.

For bug reports, include the model filename/hash, runtime, device and preprocessing steps. Share non-sensitive reproductions, and use validation data for tuning rather than repeatedly selecting changes against the published test set. Open an issue or submit a pull request.