arvinsingh/ddpm-cifar10

DDPM trained on CIFAR-10

A Denoising Diffusion Probabilistic Model (DDPM) trained on CIFAR-10 using distributed training across 6 V100 GPUs.

Model Description

• Architecture - U-Net with attention (128 base channels)
• Training - 100 epochs on CIFAR-10 (50,000 images)
• Hardware - 3 nodes × 2 V100-16GB GPUs
• Framework - PyTorch DDP

Training Details

Parameter	Value
Batch Size	64 per GPU (384 effective)
Learning Rate	2e-4
Timesteps	1000
EMA Decay	0.9999

Usage

import torch
from models import UNet, GaussianDiffusion

model = UNet(in_channels=3, out_channels=3, base_channels=128, 
             channel_mults=(1,2,2,2), num_res_blocks=2, attention_resolutions=(2,))
model.load_state_dict(torch.load("model_ema.pt"))
model.eval()

diffusion = GaussianDiffusion(timesteps=1000)
samples = diffusion.sample(model, image_size=32, batch_size=16, channels=3)

Training Code

GitHub Repository

Citation

@misc{darkbird2026,
  author = {Arvin Singh},
  title = {Darkbird: Distributed Training Examples},
  year = {2026},
  url = {https://github.com/arvinsingh/Darkbird}
}