Stable Diffusion Settings and Storing Your Images
In this article, we explore the impact of different settings used for the Stable Diffusion model and how you can store your generated images for quick reference.
Created on August 22|Last edited on January 17
Comment
This article will explore the impact of different settings used for the newly released Stable Diffusion model. It will also show you how you can store your generated images for quick reference and download later.
If you would like to replicate the Weights & Biases Tables logged in this article you can use this colab.
Big Image Bang
StableDiffusion, created by StabilityAI, arrived with a bang. The model weights were released to immediate enthusiasm. With the promise of "1 trillion images in your pocket," over 100,000 people began using the accompanying DreamStudio online service within just a few hours of the public release.
But what are the best settings for Stable Diffusion? And how do we get the most out of this newly released model?
Let's find out.
Stable Diffusion Settings
The StableDiffusionPipeline from the HuggingFace Diffusers library comes with 2 main parameters to set, the guidance_scale and the num_inference_steps.
- num_inference_steps: The number of inference steps, a higher value will produce better results
- guidance_scale: How strictly the model should adhere to the text prompt given
What value of guidance_scale or num_inference_steps should we use?
That is the question we're going to answer through a number qualitative explorations of the image generations, all done using Weights & Biases Tables.
Baseline Image Generation Benchmarking
To start, we need a baseline. Here are the 5 prompts we'll be benchmarking our exploration against:
- a photograph of car from the future, black background, black and white, ominous
- a cat in a hat
- a steampunk killer seal
- a portrait of a person dressed up as a humpback whale, victorian art
- a happy cartoon puppy dancing in the jungle
Below you will see a prompt, 8 generations, and the guidance_scale and num_inference_steps used (you may have to scroll right just a bit to see those).
These are pretty much the defaults for the pipeline, so it seems like a good place to start. Over the next few experiments, we will assess how the quality of these image generations changes as we adjust the settings.
Run: drawn-bird-36
1
Work in Progress
Add a comment
Iterate on AI agents and models faster. Try Weights & Biases today.