Overview & Model Architecture
This model repository hosts a multi-modal vision-language feature extractor based on OpenAI's CLIP architecture, optimized for digital wardrobe management and intelligent outfit recommendation pipelines.
By mapping garment photographs and aesthetic mood queries into a shared 512-dimensional latent vector space, the system computes visual compatibility scores and cosine distances between clothing items (Top, Bottom, Shoes, and Accessories).
Key Features
- 512-Dimensional Visual Style Embeddings: Extracts compact vector representations for garment images to build personal digital closets.
- Multi-Modal Matching: Matches visual anchor pieces (e.g., a specific jacket) with text-based mood or event contexts (e.g., "Casual Urban Chic", "Night Out").
- Polyvore & Fashion Compatibility: Trained/evaluated using compatibility structures inspired by the Polyvore Outfits and DeepFashion benchmarks.
- Continuous Active Learning: Designed to ingest real-time user interaction logs (
like / dislike feedback pairs) to dynamically adapt recommendation weights to personal style preferences.
license: mit
datasets:
- owj0421/polyvore-outfits
- mvasil/polyvore-outfits
- sahirp/deepfashion2
- lirus18/deepfashion
language:
- en
- tr
metrics:
- accuracy
- precision
- recall
- f1
- mean_iou
base_model:
- openai/clip-vit-base-patch32
pipeline_tag: zero-shot-classification
library_name: transformers
tags:
- vision
- clip
- fashion
- style-embeedings
- outfit-recommendation
