CS 180: Project 5

Part 1:

I felt this was really straightforward just following the given instructions.

Q1.1:

Q1.2:

Q1.3:

Q1.4:

Q1.5:

Q1.6:

Q 1.7:

Q 1.71:

Web Downloaded Image:

My drawing:

Q 1.72:

Q1.73:

Q 1.8:

Q 1.9:

I honestly learnt a lot playing around with diffusion models and I found the hybrid images incredibly fascinating when I was working with them.

Part 2:

The implemented U-Net architecture is designed for image denoising tasks, where it processes noisy images to predict and remove noise components from the original image. The network's structure comprises:

• Downsampling blocks that capture global context

• Upsampling blocks that restore spatial details

• Skip connections that preserve fine-grained information This architectural design makes U-Net particularly effective for image-to-image transformation tasks, including denoising.

Below we can visualize the effects of sigma = 0.5 when we add noise to the image.

The MNIST dataset was used for training with the following hyperparameters:

• Batch size: 256

• Number of epochs: 5

• Noise level (σ): 0.5

• Optimizer: Adam

• Learning rate: 0.0001

Here we can see the differences between loss values at epoch level 1 vs epoch 5.

This image represents performance with different sigma values.

I wasn’t able to finish but honestly, I found that opportunity of implementing an actual diffusion model and research paper to be one of the more interesting things I’ve done in my time at college.