ERA_Session20

Objective:

The purpose of this repository is to understand the architecture of Generative Art & Stable Diffusion

Repository:

.
├── LICENSE
├── README.md
├── config.py
├── diffusion_loss.py
├── image_generator.py
├── inference.ipynb
├── model.py
├── prediction.py
├── requirements.txt
├── symmetry_loss_analysis.py
└── utils.py

How to execute this repository?

In inference.ipynb, - add the prompt in the prompt variable - configure the required loss function and execute the prediction function

Results

prompt = A King riding a horse

1. Without Loss Function

2. Blue Channel

Computing the average absolute difference between the blue channel values of each pixel in the batch and the target value of 0.9. This allows us to measure how far, on average the blue channel deviates from the desired value of 0.9 across all images in the batch

3. Elastic Deformations

A data augmentation process. Applying the random elastic deformations to get an input image. The Strength and Smoothness of these deformations are controlled by the alpha and sigma parameters. The process involves generating displacement vectors for each pixel, adding these vectors to an identified grid, and then using the deformed grid to interpolate pixel values from the original image.

4. Saturation

Applied a saturation adjustment to the images, and the error is calculated as the mean absolute pixel-wise difference between the original and the transformed images