Dressing the Imagination: Text-to-Fashion Outfits with NeRA Adapters

Official implementation of “Dressing the Imagination: A Dataset for AI-Powered Translation of Text into Fashion Outfits and A Novel NeRA Adapter for Enhanced Feature Adaptation”, accepted at WACV 2026.

---

✨ Highlights

• 📚 New fashion dataset for text-to-fashion outfit generation.
• 🧠 NeRA adapter for efficient feature adaptation.
• 🚀 End-to-end training and inference pipeline included.
• 🔧 Model-agnostic adapter – demonstrated with FLUX, but easily extendable to other architectures.

---

🧩 What is NeRA?

NeRA (Nonlinear low-rank Expressive Representation Adapter) is a novel parameter-efficient fine-tuning adapter inspired by Kolmogorov-Arnold Networks (KANs), replacing MLP-based transformations in methods like LoRA with learnable spline-based activations for superior modeling of complex, nonlinear semantic relationships.

🔁 Although this implementation demonstrates NeRA with the Flux model, NeRA is model-agnostic and can be integrated with any compatible architecture. The provided scripts serve as a reference and can be easily adapted to your preferred model.

---

⚙️ Installation & Setup

1. 🐍 Create a Virtual Environment

Ensure you have Python 3.10+ installed.

# Create and activate virtual environment
python -m venv venv
source venv/bin/activate         # On Windows: venv\Scripts\activate

2. 📦 Install Dependencies

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

---

🏋️‍♀️ Training with NeRA

Update the config file training/flux_nera_config.yaml with your:

• Dataset paths
• Cache and output directories
• NeRA hyperparameters

Then run:

python train_nera_flux.py --config training/flux_nera_config.yaml

📦 Outputs:

• adapter.pt (learned adapter weights)
• config.json (model & adapter config)

---

🔍 Inference

To perform inference with trained NeRA point infer_flux_Nera.py to your adapter directory (weights + config), then run:

python infer_flux_Nera.py

---

📄 YAML Configuration includes:

• Pretrained model path
• Dataset root (images + CSV)
• NeRA hyperparameters (rank, alpha, target layers)
• Training params (batch size, LR, epochs)

---

📖 Citation

If you use this work in your research, please cite our paper:

@inproceedings{Deshmukh_2026_WACV,
  author    = {Deshmukh, Gayatri and De, Somsubhra and Sehgal, Chirag and Gupta, Jishu Sen and Mittal, Sparsh},   
  title     = {Dressing the Imagination: A Dataset for AI-Powered Translation of Text into Fashion Outfits and A Novel NeRA Adapter for Enhanced Feature Adaptation},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month     = {March},
  year      = {2026},
}