I have been very interested in learning how to use AI art generation to improve the sort of work I was doing anyway and I’ve been held utterly spellbound by the way that Stable Diffusion has, since its release allowed me to give concrete form to ideas at tremendous scale and with glorious speed. As such, I wanted to document the process I used to create this Exotica Album cover just now:

First of all, I needed images of exotic women with blue hair. As it happens I had been doing this anyway to illustrate the script of a rock’n’roll musical based on Paradise Lost - so that was handy. At the time I had been using the demo here to draw little in browser sketches and then run img2img on them - after some trial and error starting with general prompts, rudimentary faces and scribbles of green to indicate where things ought to be I ended up with some good results:

Of these, the middle one was closest to the sort of thing Exotica album covers require so I upscaled it with Real-ESRGAN using face correction to get this:

I resized the upscaled image back to 512x512 and fed it back into img2img using the prompt:

”exotica album cover. mysterious beautiful woman. tiki culture. 70s. loungecore.”

I’m particularly interested in SD’s ability to generate the sort of image that would never trend on artstation: authentic kitsch, Blake etchings, uncanny recipe blog photos, 70s pulp sci-fi illustrations and folk horror woodcuts like this one of Jason Voorhees:

It seems to me that the real artistic core of this technology is its ability to make trivial the technical barriers between the conceptualisation of ideas and their presentation as artistic products. Ideas are the key ingredients - the requirement for proficiency in mechanical skill is a hindrance to their communication and should not be fetishised. Similarly, I think it’s a mistake to talk of ‘prompt engineering’ and to think in the narrow terms that the ‘trending on Artstation’ token is emblematic of. The prompt is the location of the artistic praxis and as such it is the entry point for the infectious matter of cliché. I resist it.

Anyway, I’m digressing.

I used: 50 steps 18.0 CFG and 0.57 strength so as to get results that stuck closely to the prompt and didn’t deviate far from the init image

The best of the results was:

Which went back into img2img with the modified prompt: ”exotica album cover. mysterious beautiful woman with purple hair. tiki culture. 70s. loungecore.“ and 0.60 strength so as to vary the output just a little bit more and hopefully add a little variety to the colours. From a batch of 4 this gave me the following - with the AI deciding all on its own to paint a butterfly on her face which… fine. good, actually.

I upscaled both with and without face enhancement as you can see.

I preferred the way the enhanced face dealt with the butterfly so that went off into photoshop to await fiddling.

One thing I was very curious to see was whether SD could replace the annoying process of trawling for CC0 textures to add grunginess to photoshop projects.

At first, I tried the prompt: “texture of a plain vinyl album sleeve” but used it with img2img of the previously loaded generation by accident and:

Resisting the urge to start writing music to suit each of these, I forced myself to move on and went back to using txt2img. I prompted “texture of a plain vinyl album sleeve. photoshop texture. grunge.” 50 steps 18.0 scale to get:

Which was kind of ideal. I tried upscaling the texture - but that had the unfortunate effect of making the image smooth and ungrungey and therefore useless, so I just stretched the texture in photoshop and used it like that, because that is fine.

It was very pleasant not to use the same CC0 texture from google images for the one millionth time.

I wanted to see the slight indentation of the record in the sleeve so I generated another batch of results using “texture of a plain vinyl album sleeve. photoshop texture. 60s” 50 steps 18 scale and got:

Again, ideal. I put the first texture in a layer in photoshop set to multiply with 100% opacity then fiddled with the saturation and levels until it blended nicely. I then put the second in another multiply layer with opacity at 55% and did the same. I also cropped the image in a little and used photoshop’s (petulant looking in the circumstances) context-aware fill to get rid of the artefacts in the corners. That got me to here:

Next I needed lettering for the title. I’ve always struggled to get the right effect when hand-drawing letters - even with a tablet they never look like they were painted organically. Could SD help? I scrawled this:

And then went back to img2img and ran it with the prompt: ”crimson brush strokes. exotica album title. tiki. tropical letters spelling EXOTIQUE” with a very low strength setting of .19 and 18.0 scale - I picked the best looking result and ran that through again with the same prompt but a slightly higher strength of .25.

The best of these went back through with the amended prompt: “textured crimson brush strokes. oil paint. exotica album title. tiki. tropical letters spelling EXOTIQUE” at .27 strength. I was looking for a result that added a little bit of variation to the red - I picked the one with the most; and that went back through with the same prompt and settings. This got me to:

This went into the upscaler, then I cut it out using a few passes of Photoshop’s ‘Object Selection’ tool and pasted it into the main project. I added my artist’s name in some classic capitalised helvetica Italics for extra smooth vibes.

Classic LPs have logos on them. I have logos I could have stuck on - but I wanted to leave as much to SD as possible, so I generated some images using: “stereo logo. old fashioned logo. 60s. LP” which gave me (among some less immediately usable options) this:

Yeah, fine. That went in.

I then tried again to get one of those logos that indicates the LP is recorded in stereo that you used to get by feeding one of the initial results into img2img with the same prompt. I iterated the best results a few times until I ended up with:

Given the size it would be on the final piece I was fine with the artefacts from the stock photo site images that SD must have been trained on (though I doubt they are, lol).

I cut off the top and bottom, stretched it vertically a bit, added some white to the lower portion and added the word ‘STEREO’ and that was done.

From there it was just a case of merging the layers and tweaking the color balance a bit and that was it - done. Here she is again:

Now I’ll have to make a new exotica record to go with it.

I can’t begin to express how many boring steps this has taken out of the realisation of this idea. I have what is now called ‘Aphantasia’ - an inability to make pictures inside my head. Perhaps this is why it feels so miraculous to be able to create these images at almost the speed with with which I can describe the desire to see them.

The speed of events since Stable Diffusion was released has felt like a kind of rehearsal for the singularity. This is how I’ve imagined it: an exponential acceleration of progress. Suddenly everything is tumbling over itself as though we were traveling toward it at lightspeed and I want to help it go even faster.

Everywhere I’ve seen the question asked - how do we help? - the answer has been ‘make guides’ - so that was the point of this. To walk through an example of SD being used as a tool to achieve a practical end. I hope that it gave you some ideas.

SI

I have been very interested in learning how to use AI art generation to improve the sort of work I was doing anyway and I’ve been held utterly spellbound by the way that Stable Diffusion has, since its release allowed me to give concrete form to ideas at tremendous scale and with glorious speed. As such, I wanted to document the process I used to create this Exotica Album cover just now:

First of all, I needed images of exotic women with blue hair. As it happens I had been doing this anyway to illustrate the script of a rock’n’roll musical based on Paradise Lost - so that was handy. At the time I had been using the demo here to draw little in browser sketches and then run img2img on them - after some trial and error starting with general prompts, rudimentary faces and scribbles of green to indicate where things ought to be I ended up with some good results:

Of these, the middle one was closest to the sort of thing Exotica album covers require so I upscaled it with Real-ESRGAN using face correction to get this:

I resized the upscaled image back to 512x512 and fed it back into img2img using the prompt:

”exotica album cover. mysterious beautiful woman. tiki culture. 70s. loungecore.”

I’m particularly interested in SD’s ability to generate the sort of image that would never trend on artstation: authentic kitsch, Blake etchings, uncanny recipe blog photos, 70s pulp sci-fi illustrations and folk horror woodcuts like this one of Jason Voorhees:

It seems to me that the real artistic core of this technology is its ability to make trivial the technical barriers between the conceptualisation of ideas and their presentation as artistic products. Ideas are the key ingredients - the requirement for proficiency in mechanical skill is a hindrance to their communication and should not be fetishised. Similarly, I think it’s a mistake to talk of ‘prompt engineering’ and to think in the narrow terms that the ‘trending on Artstation’ token is emblematic of. The prompt is the location of the artistic praxis and as such it is the entry point for the infectious matter of cliché. I resist it.

Anyway, I’m digressing.

I used: 50 steps 18.0 CFG and 0.57 strength so as to get results that stuck closely to the prompt and didn’t deviate far from the init image

The best of the results was:

Which went back into img2img with the modified prompt: ”exotica album cover. mysterious beautiful woman with purple hair. tiki culture. 70s. loungecore.“ and 0.60 strength so as to vary the output just a little bit more and hopefully add a little variety to the colours. From a batch of 4 this gave me the following - with the AI deciding all on its own to paint a butterfly on her face which… fine. good, actually.

I upscaled both with and without face enhancement as you can see.

I preferred the way the enhanced face dealt with the butterfly so that went off into photoshop to await fiddling.

One thing I was very curious to see was whether SD could replace the annoying process of trawling for CC0 textures to add grunginess to photoshop projects.

At first, I tried the prompt: “texture of a plain vinyl album sleeve” but used it with img2img of the previously loaded generation by accident and:

Resisting the urge to start writing music to suit each of these, I forced myself to move on and went back to using txt2img. I prompted “texture of a plain vinyl album sleeve. photoshop texture. grunge.” 50 steps 18.0 scale to get:

Which was kind of ideal. I tried upscaling the texture - but that had the unfortunate effect of making the image smooth and ungrungey and therefore useless, so I just stretched the texture in photoshop and used it like that, because that is fine.

It was very pleasant not to use the same CC0 texture from google images for the one millionth time.

I wanted to see the slight indentation of the record in the sleeve so I generated another batch of results using “texture of a plain vinyl album sleeve. photoshop texture. 60s” 50 steps 18 scale and got:

Again, ideal. I put the first texture in a layer in photoshop set to multiply with 100% opacity then fiddled with the saturation and levels until it blended nicely. I then put the second in another multiply layer with opacity at 55% and did the same. I also cropped the image in a little and used photoshop’s (petulant looking in the circumstances) context-aware fill to get rid of the artefacts in the corners. That got me to here:

Next I needed lettering for the title. I’ve always struggled to get the right effect when hand-drawing letters - even with a tablet they never look like they were painted organically. Could SD help? I scrawled this:

And then went back to img2img and ran it with the prompt: ”crimson brush strokes. exotica album title. tiki. tropical letters spelling EXOTIQUE” with a very low strength setting of .19 and 18.0 scale - I picked the best looking result and ran that through again with the same prompt but a slightly higher strength of .25.

The best of these went back through with the amended prompt: “textured crimson brush strokes. oil paint. exotica album title. tiki. tropical letters spelling EXOTIQUE” at .27 strength. I was looking for a result that added a little bit of variation to the red - I picked the one with the most; and that went back through with the same prompt and settings. This got me to:

This went into the upscaler, then I cut it out using a few passes of Photoshop’s ‘Object Selection’ tool and pasted it into the main project. I added my artist’s name in some classic capitalised helvetica Italics for extra smooth vibes.

Classic LPs have logos on them. I have logos I could have stuck on - but I wanted to leave as much to SD as possible, so I generated some images using: “stereo logo. old fashioned logo. 60s. LP” which gave me (among some less immediately usable options) this:

Yeah, fine. That went in.

I then tried again to get one of those logos that indicates the LP is recorded in stereo that you used to get by feeding one of the initial results into img2img with the same prompt. I iterated the best results a few times until I ended up with:

Given the size it would be on the final piece I was fine with the artefacts from the stock photo site images that SD must have been trained on (though I doubt they are, lol).

I cut off the top and bottom, stretched it vertically a bit, added some white to the lower portion and added the word ‘STEREO’ and that was done.

From there it was just a case of merging the layers and tweaking the color balance a bit and that was it - done. Here she is again:

Now I’ll have to make a new exotica record to go with it.

I can’t begin to express how many boring steps this has taken out of the realisation of this idea. I have what is now called ‘Aphantasia’ - an inability to make pictures inside my head. Perhaps this is why it feels so miraculous to be able to create these images at almost the speed with with which I can describe the desire to see them.

The speed of events since Stable Diffusion was released has felt like a kind of rehearsal for the singularity. This is how I’ve imagined it: an exponential acceleration of progress. Suddenly everything is tumbling over itself as though we were traveling toward it at lightspeed and I want to help it go even faster.

Everywhere I’ve seen the question asked - how do we help? - the answer has been ‘make guides’ - so that was the point of this. To walk through an example of SD being used as a tool to achieve a practical end. I hope that it gave you some ideas.

SI