Stylique/dressify-models

Dressify Outfit Recommendation Models

This repository contains the trained models for the Dressify outfit recommendation system.

Models

ResNet Item Embedder

• Architecture: ResNet50 with custom projection head
• Purpose: Generate 512-dimensional embeddings for fashion items
• Training: Triplet loss with semi-hard negative mining
• Input: Fashion item images (224x224)
• Output: L2-normalized 512D embeddings

ViT Outfit Compatibility Model

• Architecture: Vision Transformer encoder
• Purpose: Score outfit compatibility from item embeddings
• Training: Triplet loss with cosine distance
• Input: Variable-length sequence of item embeddings
• Output: Compatibility score (0-1)

Usage

from huggingface_hub import hf_hub_download
import torch

# Download models
resnet_path = hf_hub_download(repo_id="Stylique/dressify-models", filename="resnet_item_embedder_best.pth")
vit_path = hf_hub_download(repo_id="Stylique/dressify-models", filename="vit_outfit_model_best.pth")

# Load models
resnet_model = torch.load(resnet_path)
vit_model = torch.load(vit_path)

Training Details

• Dataset: Polyvore Outfits (Stylique/Polyvore)
• Loss: Triplet margin loss
• Optimizer: AdamW
• Mixed Precision: Enabled
• Hardware: NVIDIA GPU with CUDA

Performance

• ResNet: ~25M parameters, fast inference
• ViT: ~12M parameters, efficient outfit scoring
• Memory: Optimized for deployment on Hugging Face Spaces

Citation

If you use these models in your research, please cite:

@misc{dressify2024,
  title={Dressify: Deep Learning for Fashion Outfit Recommendation},
  author={Stylique},
  year={2024},
  url={https://huggingface.co/Stylique/dressify-models}
}