Model Card for ddpm-unet-retinal-fundus-image-generator

A U-Net–based Denoising Diffusion Probabilistic Model (DDPM) trained to generate retinal fundus images. This model can be used for synthetic medical image generation to augment datasets for training diagnostic models or other biomedical tasks.

---

Model Architecture

• Base: UNet2DModel
• Scheduler: DDPMScheduler
• Resolution: 128x128
• Channels: RGB (3)
• Attention: Spatial self-attention in mid-resolution blocks
• Framework: 🤗 Diffusers + PyTorch

---

Dataset

• Source: Kaggle - Retinal Fundus Images
• Subset Used: train/Moderate Diabetic Retinopathy
• Preprocessing:
  • Resized to 128x128
  • Normalized to [-1, 1]
  • Random horizontal flip

---

Training Configuration

Setting	Value
Epochs	35
Batch size	16
Optimizer	AdamW
Learning rate	1e-4
Scheduler	Cosine w/ warmup
Precision	Mixed (fp16)
Diffusion Timesteps	1000
Image Samples Saved	Every 10 epochs

Training was done using 🤗 Accelerate and TensorBoard logging.

---

How to Use

from diffusers import DDPMPipeline
import torch
import matplotlib.pyplot as plt

pipeline = DDPMPipeline.from_pretrained("GS-23/ddpm-unet-retinal-fundus-image-generator")
images = pipeline(batch_size=1, generator=torch.manual_seed(0)).images

for img in images:
    plt.imshow(img)
    plt.axis("off")
    plt.show()

---

📌 Use Cases

• Data augmentation for diabetic retinopathy classifiers
• Retinal pathology simulation and training
• Medical generative AI research
• Domain-specific image synthesis