Diffusion-CIFAR10

Diffusion-CIFAR10 — Unconditional DDPM on CIFAR-10

A Denoising Diffusion Probabilistic Model (DDPM) trained on CIFAR-10 for unconditional image generation. The model generates 32x32 RGB images spanning all 10 CIFAR-10 categories without class conditioning.

Model Description

Implements DDPM (Ho et al., 2020) with a U-Net denoising network featuring self-attention blocks and sinusoidal time embeddings.

Architecture — U-Net

Time Embedding:

• Sinusoidal timestep encoding → two-layer FFN with Swish activation
• Injected as an additive bias into every residual block

Residual Blocks (ResBlock):

• Two 3x3 Conv2d with GroupNorm and Swish activation
• Time embedding added between convolutions
• 1x1 projection on skip path when input/output channels differ

Attention (AttnBlock):

• Single-head self-attention (query, key, value projections)
• Applied at the bottleneck and selected resolution levels

Downsampling: Strided Conv2d (stride 2) Upsampling: Nearest-neighbour interpolation + Conv2d

Diffusion Process

• Parameterization: Noise prediction (epsilon)
• Schedule: Linear beta schedule
• Loss: MSE between predicted and actual noise

Training Details

Repository Contents

How to Use

import torch
from model import UNet
from diffusion import GaussianDiffusionSampler

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = UNet(T=1000, ch=128, ch_mult=[1,2,2,2], attn=[1], num_res_blocks=2, dropout=0.1)
model.load_state_dict(torch.load("Checkpoints/ckpt_199.pth", map_location=device))
model.eval().to(device)

sampler = GaussianDiffusionSampler(beta_1=1e-4, beta_t=0.02, model=model, T=1000).to(device)

with torch.no_grad():
    x_T = torch.randn(16, 3, 32, 32, device=device)
    samples = sampler(x_T)

samples = (samples.clamp(-1, 1) + 1) / 2  # rescale to [0, 1]

References

• Ho et al. (2020). Denoising Diffusion Probabilistic Models

License

MIT