A tale of a Coyote and a possum in a backyard ramble
All right, this story is as fake as its supposed protagonists. Or rather, they are made up like the story 😆. It is true that with the advent of Stable diffusion from STABILITY AI it is now possible to create large quantities of images for all kinds of creatures. Even those that might haunt your backyard day or night 🎃.
In this case, we are not invoking any fantasy spell but instead seeking to create “good enough” images of animals to use for training AI models.
For our purpose, AUTOMATIC1111 stable-diffusion-webui provides a browser interface for Stable diffusion.
Our friends at ONESTEPAI kindly let us you one of their vGPU to run it.
In the beginning, the text prompts we used were simple, such as “coyote in a backyard” or “squirrel on a fence”.
We used the Stable diffusion version v1.4 with the following settings:
It took around 10 seconds per 512x512 picture. We can batch-process our text prompts to create an arbitrary number of images. You can add text under the negative prompt to further guide the creation process.
At the end of this initial attempt, we gained enough confidence that the image datasets, while far from perfect, might be good enough. A future blog will explore training a simple AI image classifier using these synthetic image datasets
Also, we will explore the latest version of Stable diffusion v1.5, including some additional options.
Stay tuned, and until then … happy trails!