MODA_NER(V) - Crop
Fix, 2026-09-19. If you ran this model before this date, re-run it. The repository's
models/inference.pyran the crop route with the wrong image preprocessing: OpenAI-CLIP normalisation and a centre crop, instead of the mean/std 0.5 squash this checkpoint was trained with. Its outputs did not match the published benchmark predictions, matching on 16 of 64 held-out images. The weights in this repository are unchanged and correct. Pull the latest Moda_ner code and the Usage command below produces the published predictions. The benchmark numbers on this card were produced correctly.
Tier * - open code, open weights.
Tier * — open code, open weights. Weights: MIT.
Conditional attribute heads on our own frozen encoder (see Provenance below).
Input contract: one localized garment crop. Output: 15 sparse fields — master category, category, sub-category, silhouette, hemline, sleeve length and shape, neckline, collar presence and style, waist, material, surface treatment, pattern, closure. No colour, no fit; the evaluation corpus for this track carries no equivalent labels.
| Attribute micro-F1 (crop track) | |
|---|---|
| This released checkpoint | 0.6300 |
| Same architecture on a frozen FashionSigLIP encoder | 0.6245 |
| Zero-shot open VLM reference (Qwen3-VL-8B) | 0.1805 |
| Zero-shot FashionSigLIP text prototypes | 0.1817 |
These are the weights that produced our published number: nothing is held back on this route. Field-macro F1 is 0.6074, category accuracy 0.8825, master-category accuracy 0.9215.
Weak fields, stated because the micro average hides them: material 0.4148, plus surface treatment, collar style, neckline and waist type. A headline of 0.6300 does not mean every field is usable.
Attribution. The annotations behind this track are licensed CC BY 4.0 and require credit: Jia et al., Fashionpedia: Ontology, Segmentation, and an Attribute Localization Dataset, ECCV 2020. No images from that corpus are redistributed in the weights or the repository.
Provenance
The encoder these heads run on is ours: HopitAI/moda-fashion-distilled, MIT, already public. Nothing from another vendor is loaded at inference time.
That is worth stating plainly, because the comparator on this track is a FashionSigLIP-based system and it would be easy to assume this model is that system with heads attached. It is not. FashionSigLIP appears in two other roles:
- As the distillation teacher. An earlier ladder of checkpoints put conditional heads on frozen Marqo-FashionSigLIP. We distilled that system into our own encoder; the teacher is used during training and is not needed to serve.
- As the baseline we measure against. The comparator figure quoted above is that same FashionSigLIP-based system.
Lineage, stated once rather than implied: moda-fashion-distilled is itself a distilled student built on ViT-B/16-SigLIP, from a teacher ensemble that included our own DeepFashion2 fine-tune. Marqo-FashionSigLIP is Apache-2.0. The DeepFashion2 corpus is research-only, so we do not describe this pipeline as provenance-clean end to end.
Credit for this model. If you use these weights or report numbers from this track, cite
the MODA General Attribute Suite (see CITATION.cff). Scorers stamp the suite name and
version into every result file for exactly this reason.
Links
- Benchmark tables and protocol: https://hopit-ai.github.io/Moda_ner/
- Code, scorers and prediction files: https://github.com/hopit-ai/Moda_ner
- All Hopit AI benchmarks: https://hopit-ai.github.io/
Usage
The heads are not a transformers architecture, so load them through the suite
repository rather than AutoModel:
git clone https://github.com/hopit-ai/Moda_ner && cd Moda_ner
pip install -r requirements-inference.txt
huggingface-cli download HopitAI/moda-ner-v-crop --local-dir ./moda-ner-v-crop
python models/inference.py --route crop --model-dir ./moda-ner-v-crop --images photo.jpg