MODA
An open benchmark, harness and set of embedding models for fashion product retrieval, letting a shopper find products from a photo or a plain-language description and measuring results at full catalog scale against other embedding models.
MODA evaluates fashion image-to-image and text-to-image retrieval by running a complete pipeline, combining BM25, SPLADE, dense retrieval, hybrid fusion, named-entity recognition and cross-encoder reranking, against 253,685 purchase-grounded queries synthesized from real H&M purchase data. Results are reported at full corpus scale with no gallery subsampling, using identical preprocessing for every model compared, including cases where the project's own models do not win, and the project reports its smaller open model as the top scorer on the LookBench evaluation for image-to-image retrieval.
The open models are distributed as open_clip-compatible checkpoints on Hugging Face; a query is encoded as a single 768-dimensional vector, compared by cosine similarity, and can be stored in any vector database without a required reranker or prompt template. A larger, hosted model, MODA Pro, is available alongside the open MODA and MODA Pro Lite models.
It is written in Python, installs with open_clip_torch, torch and pillow, and is released under the MIT license.