Diffusion-Based Image Super-Resolution

A residual-diffusion model for 4x image super-resolution on DIV2K.

A diffusion model for 4× image super-resolution on DIV2K, trained on an NVIDIA A100 GPU.

  • Architecture: time-conditioned U-Net with Swin Transformer attention.
  • Diffusion process: 15-step residual diffusion for reconstructed high-resolution images.
  • Training: mixed-precision execution with Weights & Biases tracking for loss, validation metrics, and reconstructed outputs.
  • Validation: 22.46 dB PSNR, 0.8709 SSIM, and 0.3907 LPIPS on the DIV2K validation set.

GitHub