---
license: apache-2.0
pipeline_tag: image-to-image
tags:
- Diffusion
- Augmentation
- PromptControlledDiffusion
- semanticsegmentation
- synthetic-data
- remote-sensing
---

<div align="center">

# 🎨 SyntheticGen

### Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation

*Addressing class imbalance in remote sensing datasets through controlled synthetic generation*

[![Accepted at IEEE IGARSS 2026](https://img.shields.io/badge/Accepted-IEEE%20IGARSS%202026-1f77b4)](#)
[![arXiv Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://huggingface.co/papers/2602.04749)
[![GitHub Code](https://img.shields.io/badge/Code-GitHub-blue?logo=github)](https://github.com/Buddhi19/SyntheticGen)
[![Live Demo](https://img.shields.io/badge/Live%20Demo-Colab-orange?logo=googlecolab&logoColor=white)](https://colab.research.google.com/drive/11KqBQogdIjwC6UXAGVeD4cfq_VclUC_I?usp=sharing)
[![Dataset](https://img.shields.io/badge/Datasets-Hugging%20Face-yellow?logo=huggingface&logoColor=black)](https://huggingface.co/datasets/buddhi19/SyntheticGenV5)

</div>

---

## 🌟 Overview

**SyntheticGen** is the official implementation for the paper [Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation](https://huggingface.co/papers/2602.04749). It tackles the long-tail distribution problem in remote-sensing datasets (specifically LoveDA) by generating synthetic imagery with *explicit control* over class ratios.

### ✨ Highlights
- **Two-stage pipeline**: Ratio-conditioned layout D3PM + ControlNet image synthesis.
- **Controllable Augmentation**: Specify exact proportions of each land cover class (e.g., `building:0.4`).
- **Data-Centric Strategy**: Improves segmentation performance by adding the *right* samples to the training set.

<div align="center">
  <img src="https://raw.githubusercontent.com/Buddhi19/SyntheticGen/main/docs/results.png" alt="SyntheticGen Results" width="100%">
</div>

---

## 🚀 Quick Start

### Installation
```bash
git clone https://github.com/Buddhi19/SyntheticGen.git
cd SyntheticGen
pip install -r requirements.txt
```

### Generate Your First Synthetic Image
To generate a synthetic image-label pair using a specific configuration:
```bash
python src/scripts/sample_pair.py \
  --config configs/sample_pair_ckpt40000_building0.4.yaml
```

---

## 📚 Usage

### Training Pipeline

**Stage A: Train Layout Generator (D3PM)**
```bash
python src/scripts/train_layout_d3pm.py \
  --config configs/train_layout_d3pm_masked_sparse_80k.yaml
```

**Stage B: Train Image Generator (ControlNet)**
```bash
python src/scripts/train_controlnet_ratio.py \
  --config configs/train_controlnet_ratio_loveda_1024.yaml
```

### Inference / Sampling
**Override config parameters via CLI:**
```bash
python src/scripts/sample_pair.py \
  --config configs/sample_pair_ckpt40000_building0.4.yaml \
  --ratios "building:0.4,forest:0.3" \
  --save_dir outputs/custom_generation
```

---

## 📄 Citation
```bibtex
@misc{wijenayake2026mitigating,
      title={Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation}, 
      author={Buddhi Wijenayake and Nichula Wasalathilake and Roshan Godaliyadda and Vijitha Herath and Parakrama Ekanayake and Vishal M. Patel},
      year={2026},
      eprint={2602.04749},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.04749}, 
}
```

---

## 🙏 Acknowledgments
- LoveDA dataset creators for high-quality annotated remote sensing data.
- Hugging Face Diffusers for diffusion model infrastructure.
- ControlNet authors for controllable generation.