![A line drawing with the caption 'GARTER: someone who advertises or repairs dots in the sea. The drawing depicts a boat floating in the ocean next to a large disc, with a person on the boat holding a ladder descending to the disc. Other discs float in the water in the distance.](/images/stablediffusion/garter-original.png)
![A version of the first image without the caption in which the outlines have been crudely coloured in with a graphics editor. Choppy waves and four fluffy white clouds have been added.](/images/stablediffusion/garter-colourised.png)
![An image generated by the StableDiffusion system using the second image as a prompt. The image now has a pleasing brushtstroke texture and the boat seems to be a speedboat with two indistinct figures.](/images/stablediffusion/garter-sd-1.png)
![An image generated by the StableDiffusion system using a darkened and desaturated version of the second image as a prompt. This one is moodier and the red disc has a kind of yellow ruffled skirt and a yellow boss in the middle. The human figure in the boat is considerably larger and better defined, although their face is blurry](/images/stablediffusion/garter-sd-2.png)
Ranjit commented on my drawing of Glossatory's definition of GARTER: "I find this one particularly evocative! Please do an oil painting next" so I colourised the original and fed it to StableDiffusion with different degrees of desaturation, and the prompt text "A man in a green boat holds a ladder to a red disc floating in a choppy sea under a blue sky with white fluffy clouds, yellow, green and purple discs in the distance, dramatic oil painting, winslow homer". I think the algorithm did pretty well.
I didn't post as many links this week because I was preoccupied helping to run the RSE Asia Australia Unconference - RSE referring to research software engineer, which is an attempt to carve out a professional identity which bridges people like myself, technical specialists who support research systems, and researchers who write and maintain software. It went really well and I plan on writing a longer post here in the next week or so.
It wasn't one of my daily links but this post about prompt injection attacks against GPT-3 is essential reading if you're interested in either machine learning or software engineering, and a great example of how the industry commits the same mistakes over and over again in different contexts.