Stable Diffusion photomedia explorations: portraiture
I've been working with Stable Diffusion since the first day it released to the public, and have recently begun teaching it to creative art teams. While Midjourney has kind of dominated the discourse online, due to it's relative ease of use and low barrier to generating great results, I've been drawn to Stable Diffusion because of the sheer amount of control it affords the user. I'm fascinated by its ability to pick up on those kind of ambient qualities that makes a photograph look photographic, so whenever I get a spare moment I've just sort of obsessively been trying to see how far it can be pushed towards producing the kinds of photographs I loved when I first stepped into the darkroom.
Naturalistic, rather than photorealistic.
There's been a lot of 'photorealistic' work generated by Midjourney, but I find most of it looks quite static in terms of pose, and more 'CGI-photorealistic' as opposed to 'lifelike'. My explorations in Stable Diffusion have been focused almost entirely on it's capabilities to render more delicate, more nuanced pose and expression within the realm of naturalistic optical images.
Posted here then are the finished* versions of the best selects from a quick Stable Diffusion session the other week. By the end I had around 100 useable images, which I narrowed down to the best 20. Tried to keep a theme for this post so I culled them down to these 10 who all look like they shared the same hairstylist.
All of these were built from rough sketches using img2img and two custom models. It's the models doing most of the heavy lifting here. The finishing process is Photoshop intensive, consisting of a lot of back and forth between generations/iterations and compositing/retouching. I was doing a lot more retouching in my earlier tests because I was using the base models and they needed a lot of error correction. Using my own custom models has cut down a lot on that, so most just need compositing now, along with some minor retouching. They might not even 'need' it, as they're probably good enough straight out of SD most of the time, but it's just the way I work.
Img2img helps guide the poses to a certain degree, but I prefer leaving SD with a little room to move on its own terms. To that end, sometimes it will interpret the sketches much more loosely, and give me something different to what I was using as a base. Sometimes it looks much more interesting, and I'm fine going in the direction of these happy little accidents, so to speak.
If you're trying to build some kind of story with these tools, it's nice to know you can get relatively consistent, harmonious visuals with specific elements, while still having some contextualised variation from image to image. The outfits look like they belong in a shoot together, but they're not just the same outfit from one image to another.
The tech is very, very capable. With some caveats.
Resolution for 'photographic' images is still an issue for now, though I don't suspect it will be for long. You can upscale the more 'digital' looking art that Midjourney excels at much more naturally, but when you're trying to work with skin texture, naturalistic gradients and optical qualities like grain and variance, you'll run into problems quicker. I don't mind these kinds of artifacts, in the same way I don't mind their optical equivalents either. But others might hate them.
All in all, a pretty intriguing look into the potential of Stable Diffusion with photomedia. Considering these will likely look primitive in six months time, I guess it's a fairly fascinating glimpse into our imaging future as well.
*I say 'finished', but the upscaling capabilities of SD, and the potential for further re-mastering/additional finishing of images like this, is practiaclly a specialist field in and of itself. So that's maybe a task for another time. These are 'finished', for now at least.