Introduction
Hey there, fellow tech heads! Ready to dive deep into the world of Stable Diffusion models? Today, we’re setting the stage with a fixed set of prompts and settings to establish a “standard” for comparing different models. This post will lay out the explanation and prompts, ensuring that all comparison pages follow the same layout (makes it easier for me too: template!).
The Game Plan
First things first, let’s outline our methodology. We’re going to test each Stable Diffusion model with a consistent set of prompts and settings. Each prompt will focus on different aspects to see how the models compare, while others are there just because I like them. Here’s what we’re looking at:
- Token Acceptance: Does the model recognise and use the first word of the prompt?
- Style Bias: What artistic style does the model lean towards?
- Strengths and Weaknesses: Where does the model excel, and where does it fall short?
- Ideal Settings: Which parameters yield the best results for each model?
- CFG and Steps: Which combinations work best for different styles?
- Aspect Ratio: Can an extreme aspect ratio be used without image duplication?
By analysing these factors, we’ll get a clear picture of how each model performs under different conditions. While this is by no means an exhaustive test, it will give us a solid idea about each model.
Token Acceptance
We’ll start some prompts with a keyword to see if the model effectively incorporates it into the generated image. This will test the model’s ability to understand and prioritise the first token in the prompt.
Style Bias
Each model tends to have a certain artistic style it gravitates towards. We’ll examine the generated images to identify any noticeable style biases.
Strengths and Weaknesses
By comparing the outputs, we’ll highlight where each model shines and where it struggles. This will help us understand the ideal use cases for each model.
CFG and Steps
We’ll use different combinations of CFG scales and steps to see how they affect the quality and style of the generated images.
Aspect Ratio
Lastly, we’ll test extreme aspect ratios to see if the models can handle them without duplicating parts of the image or losing coherence.
Conclusion
By sticking to this structured approach, we can fairly and thoroughly compare different Stable Diffusion models. Keep an eye out for my upcoming comparison posts where we’ll dive into the strengths and weaknesses of each model.