bala1802's picture
Upload README.md
51ce0fb verified
|
raw
history blame
1.69 kB
# ERA_Session20
## Objective:
The purpose of this repository is to understand the architecture of Generative Art & Stable Diffusion
## Repository:
```
.
β”œβ”€β”€ LICENSE
β”œβ”€β”€ README.md
β”œβ”€β”€ config.py
β”œβ”€β”€ diffusion_loss.py
β”œβ”€β”€ image_generator.py
β”œβ”€β”€ inference.ipynb
β”œβ”€β”€ model.py
β”œβ”€β”€ prediction.py
β”œβ”€β”€ requirements.txt
β”œβ”€β”€ 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
![Alt text](image.png)
### 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
![Alt text](image-1.png)
### 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.
![Alt text](image-2.png)
### 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
![Alt text](image-3.png)