Environment Drinks

Image Generation
Fast Stable Diffusion
Dream Studio
Remini Web
Photoshop

I have always had a passion for baking and combining my favorite plants into delicious desserts. The idea of this work is to take this a step further and use the real-world environment as a source of inspiration by combining different colors and types of drinks to create a unique beverage that is representative of the natural world. With the help of Stable Diffusion (an AI image generator), I began to experiment with different cocktails and their bright colors to represent different environments, such as a Blood Mary for a volcano and a mint mojito for streams and green leaves. However, I soon realized that I was limiting myself if I only used cocktails and so I decided to explore further and create drinks that were representative of different natural environments.

I take into consideration the iconic items that are found in each environment and use them to create a visual representation of the environment. For example, I used coconut trees and undersea to show the characteristics of the beach. And I didn't only use Stable Diffusion to generate the images, I also took the result into Photoshop to create each drink poster to present these in a better way.

I do find the process of creating these drinks to be a creative challenge. Thinking about what elements should be in the enviroment and positioning all the items to make the drink looks good are the biggest challenges I was facing. The combination of the two interests allows me to create something that is unique and visually appealing. I believe that my work let people be able to appreciate the natural beauty of the environment shown in a form of a drink.

View on Github

Source

There are several software been used in this project.

The main tool to train Stable Diffusion with my own images: fast stable diffusion.
The main tool to generate images: Stable Diffusion
Tool to get the resource images for training: Dream Studio
Tool to enhance the quality/resolution of the generated image: Remini Web
Tool to edit the generated image into a poster: Photoshop

Settings in Stable Diffusion

txt2img: Sampling method: Euler a Sampling steps: 150 Width: 512 Height: 512 Others: Default
img2img: Inpaint Mask blur: 4 Mask mode: Inpaint masked Masked content: latent noise Inpaint area: Only masked padding, pixels: 32 Sampling steps: 150 Others: Default

Method

This project is done in several platforms/software. The steps will be listed in numbers:

Before generating any images, training the AI image generator (Stable Diffusion) with special keywords and images is the very first step. I want my Stable Diffusion to learn everything about the enviroment and stuff I want to show in the enviroment. So I used fast stable diffusion to train my own image generator.
I used Dream Studio as my resource image finder. Instead of finding low-resolution images on Google, I chose Dream Studio to actually let it generate the images with the prompts I provide to get all in one-size (512 x 512) and high-resolution images. For example, I want all my drinks and resources to be isometric, I used isometric, mojito, pixel art, and hyper details as my prompt to generate some images I can use to train my own stable diffusion. So it's kind of interesting that I'm using a result from an AI image generator to train another AI image generator.
After I got all my images, I have to rename the same type into the same names (the keyword I will later use to generate as prompt) with sorting numbers, for example, mojito(1), mojito(2), mojito(3) and etc.
Then I will start on my fast stable diffusion on Google Colab and create my session name (project2pocky) in Create/Load a Session section, and then upload all my images in the Instance Images section. After it uploads all the images, the training begins. I set it to 3000 UNet Training Steps, with 2e-6 Learning Rate, and 350 Text Encoder Training Steps, 1e-6 Text Encoder Learning Rate. After it's trained, I can test the trained model and start my trip. I noticed it will be really helpful if you name the images to special names that the basic model doesn't have these keywords in. When you generate the image, it will know exactly what you want.
For the generate part, I will first use simple prompts to generate from text to image, and then put the image I feel it's good into img2img Inpaint. I will erase some of the parts and use different prompts to put the items I want on the image. I spent most of my time on this step. Usually, an image took about 1-2 hours to get a result I like. I did download about 200 images to make 10 static images.
After I download the result I like, I will put it on Remini Web to have the image in a higher resolution.
Then I will put the image into Photoshop and cut off the background of the image. And use the colors from the image to make a poster that presents the drinks in a better way.

Rose Bush - Coke

Beach - Sparkling Water

Beach text2img: isometric glass cup, isometric sparkling water in green blue img2img Inpaint: beach pixel art with sea; coconut tree isometric; underwater pixel art