Replace Gaussian Noise with Gaussian Blur for Forward Diffusion

Description

In our current implementation of the forward process in the DDPM (Denoising Diffusion Probabilistic Model), we use Gaussian noise to iteratively corrupt the image. While Gaussian noise works well, we want to explore using Gaussian blur instead, which may provide different inductive biases and could be useful for tasks where structural preservation in early timesteps is beneficial.

Gaussian blur has been explored in the context of diffusion models and generative processes. It introduces a different type of degradation by smoothing local details instead of adding pixel-wise noise, potentially offering advantages in image synthesis and restoration tasks.

Tasks

• Replace Gaussian noise with Gaussian blur in the forward diffusion process.
• Experiment with different kernel sizes and sigma values for the blur.
• Analyze whether the learned reverse process can effectively deblur images to reconstruct clean samples.
• Compare training and sampling results between noise-based and blur-based diffusion.

Relevant Papers & Resources

• Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise – Proposes diffusion models where noise is replaced with other degradation functions, including blur.
• BLURRING DIFFUSION MODEL - Describes a new type of diffusion process for generative modeling based on heat dissipation, or blurring, as an alternative to isotropic Gaussian diffusion.

Why This Matters

Exploring blur as an alternative to noise in the forward process may open new possibilities for training diffusion models that are more robust or better suited for specific tasks like deblurring, super-resolution, and artistic style transfer. If this works well, we could even generalize the forward process to other types of degradations beyond just noise.

Would love to hear thoughts and contributions from the community! 🚀

No file, test, or notebook entry point is named in the issue. First locate the forward diffusion implementation in the repository, then compare the existing noise-based process with blur settings across training and sampling; done means reporting whether clean samples can be reconstructed and how results compare.