Live app: https://lakshyalol-fake-dataset-factory.hf.space
Generate labeled synthetic chest X-rays using 6 generative architectures spanning 2014–2022. Evaluate each model individually. Export a usable dataset. Compare them honestly.
Real chest X-ray data is expensive, private, and class-imbalanced.
This tool generates synthetic labeled chest X-ray images using multiple generative AI approaches — and then actually measures if the synthetic data is any good.
Input: pick a model, pick a class (Normal / Pneumonia), pick how many images
Output: images/ + labels.csv + metrics.json — ready to plug into any classifier
| Model | Year | What Makes It Different |
|---|---|---|
| Stable Diffusion | 2022 | prompt-based img2img, no training needed |
| DCGAN | 2014 | baseline GAN, BCE loss |
| WGAN-GP | 2017 | stable GAN, Wasserstein + gradient penalty |
| VQ-VAE | 2017 | discrete codebook, sharper than vanilla VAE |
| DDPM | 2020 | diffusion model, strong quality |
| Flow Matching v2 | 2022 | best quality — attention UNet + EMA + adaptive ODE |
Each model trained on the same dataset. Same evaluation. Apples to apples (except SD — documented).
| Model | Year | FID ↓ | TSTR ↑ | Notes |
|---|---|---|---|---|
| Flow Matching v2 | 2022 | 6.20 | 100% | 🥇 Best — attention + EMA + 100 epochs |
| DDPM | 2020 | 8.96 | 100% | 🥈 Diffusion baseline |
| WGAN-GP | 2017 | 11.10 | 100% | 🥉 Stable GAN training |
| DCGAN | 2014 | 15.24 | 100% | Strong baseline |
| VQ-VAE | 2017 | 46.59 | 100% | No PixelCNN prior* |
| Stable Diffusion | 2022 | 94.71 | 100% | img2img mode** |
* VQ-VAE uses random codebook sampling — proper generation requires PixelCNN prior on learned indices.
** SD img2img uses real images as input, not pure noise. High FID reflects different input distribution by design.
Domain-adapted FID — DenseNet121 features from torchxrayvision instead of Inception V3 (trained on ImageNet). Medically meaningful comparison.
Proxy TSTR — generated images scored by pretrained torchxrayvision classifier. Lung Opacity > 0.5 = synthetic image carries medically plausible features. Lung Opacity used over strict Pneumonia label because generative models learn visual patterns, not clinical diagnoses.
import torchxrayvision as xrv
model = xrv.models.DenseNet(weights="densenet121-res224-all")
# pretrained on NIH + CheXpert + MIMIC + OpenI + KaggleSource: paultimothymooney/chest-xray-pneumonia (Kaggle, free)
Preprocessed: lakshyarathi/lungpp (64×64 grayscale, ready to use)
chest_xray/
train/
NORMAL/ 1341 images
PNEUMONIA/ 3875 images
test/
NORMAL/ 234 images ← TSTR testing
PNEUMONIA/ 390 images
Local path: chest_xray/ or Kaggle: /kaggle/input/chest-xray-pneumonia/chest_xray/
fake-dataset-factory/
├── notebooks/
│ ├── 01_stable_diffusion.ipynb
│ ├── 02_dcgan.ipynb
│ ├── 03_wgan_gp.ipynb
│ ├── 04_vqvae.ipynb
│ ├── 05_ddpm.ipynb
│ ├── 06_flow_matching.ipynb
│ ├── 06_flow_matching_v2.ipynb ← improved version (best FID)
│ ├── 07_comparison.ipynb
│ └── outputs/
│ ├── 01_stable_diffusion/ ← images, labels.csv, metrics.json
│ ├── 02_dcgan/
│ ├── 03_wgan_gp/
│ ├── 04_vqvae/
│ ├── 05_ddpm/
│ ├── 06_flow_matching/
│ ├── 06_flow_matching_v2/ ← best results
│ ├── combined_metrics.json ← all models compared
│ └── comparison_table.csv
├── data/
│ ├── prepare.ipynb
│ ├── normal/ ← preprocessed 64×64
│ └── pneumonia/
├── src/
│ ├── models/
│ ├── evaluate.py
│ ├── label.py
│ └── export.py
├── app.py ← Gradio UI
└── README.md
pip install torch torchvision diffusers transformers accelerate
pip install torchcfm torchxrayvision pytorch-fid gradioGPU recommended. All notebooks run on Kaggle T4 (free, 30hr/week).
python app.pyOr run any notebook independently — each is self-contained, saves to outputs/<model_name>/.
Gradio app auto-detects which models are trained and enables them.
- WGAN-GP on chest X-rays: MDPI 2023
- VQ-VAE on CheXpert: kamenbliznashki/generative_models
- Flow Matching for medical imaging: arxiv 2503.00266 (March 2025)
- torchxrayvision: mlmed/torchxrayvision