Case Study: Sara
I wanted to make a self portrait in the style of the “Woman from Granada”
by Hermen Anglada Camarasa, a Spanish painter active in the early 1900s
whose work is reminiscent of post-impressionism and the Vienna Secession.
"Woman from Granada" by Hermen Anglada Camarasa, circa 1914
I started with DALL-E and a very simple prompt.
"woman in her 30s curly brunette green eyes Hermen Anglada i Camarasa"
The results were photographic and I suspect that,
instead of referencing the painter, it was referencing
images of people with similar names. I added “painting by”
to my prompt and the results were redirected to the painter.
"painting by Hermen Anglada i Camarasa woman in her 30s curly brunette green eyes"
The next several prompts added more and more details to get the
image to look more like me, and background details to look more
like the painting I was referencing.
"painting by Hermen Anglada i Camarasa woman in her 30s loose curly brunette with bangs green eyes square face"
"painting by Hermen Anglada i Camarasa woman in her 30s medium length loose curly brunette with bangs green eyes square face spanish shawl"
"painting by Hermen Anglada i Camarasa woman in her 30s medium long length loose curly brunette with bangs green eyes square face flowers spanish shawl nocturnal"
I tried to make the image include more of the body, but instead I got a body without head or feet.
"painting by Hermen Anglada i Camarasa woman in her 30s medium long length loose curly brunette with bangs green eyes square face flowers spanish shawl nocturnal full body colorful painterly"
I started trying to manipulate the features more. The default was to more delicate features, and I do not have a delicate nose or a little pointy chin. The images started to go further in the right direction.
"painting by Hermen Anglada i Camarasa full portrait of a woman in her 30s medium long length loose curly brunette with bangs green eyes square jaw big nose high cheekbones olive skin flowers and floral Spanish shawl nocturnal colorful impressionist"
"painting by Hermen Anglada i Camarasa full portrait of a woman in her 30s medium long length loose curly brunette with bangs green eyes square jaw big nose high cheekbones big facial features olive skin flowers and floral Spanish shawl nocturnal colorful impressionist"
"painting by Hermen Anglada i Camarasa full portrait of a woman in her 30s medium long length loose curly brunette with bangs green eyes square jaw big nose high cheekbones big facial features olive skin flowers and floral Spanish shawl nocturnal colorful impressionist"
One of these prompts produced an image that reminded me of my mother, but it wasn’t right for me.
"painting by Hermen Anglada i Camarasa full portrait of a woman in her 30s loose curly brunette with bangs green eyes square jaw big nose high cheekbones olive skin flowers and floral Spanish shawl nocturnal colorful impressionist"
I revised my facial description after a discussion with my cousin who I live with and had informed feedback about what I actually look like. This is when the images really started to look more like me.
"painting by Hermen Anglada i Camarasa full portrait of a woman in her early thirties loose curly brunette with bangs hooded big hazel green eyes square face weak jaw big nose full lips high strong cheekbones olive skin beauty marks flowers floral Spanish shawl nocturnal colorful impressionist"
"painting by Hermen Anglada i Camarasa full portrait of a woman in her early thirties loose curly brunette with bangs hooded big golden green eyes square face weak jaw big nose full lips high strong cheekbones pale olive skin beauty marks flowers floral Spanish shawl nocturnal colorful impressionist"
The last prompt ultimately had the image I went with, but I played around more with the prompt. When I tried to make the shawl drape the way that it did in the original painting, the images started to look like renaissance portraits.
"Woman from Granada painting by Hermen Anglada i Camarasa full portrait of a woman in her early thirties with her head tilted to the side holding a flower long hair curly brunette with bangs hooded big green eyes square face weak jaw big nose full lips sly smile high strong cheekbones pale glowing olive skin beauty marks torso wrapped in multicolor flower printed purple Spanish shawl bare decolletage nocturnal dark blue garden colorful impressionist"
I also tried taking out the painter’s name entirely and only using words that described his art style. The results didn’t look vastly different and were possibly even more reminiscent of the artist.
"Impressionist Spanish Modernist chromaticism Spanish decorative style nocturnal interior subject Vienna Secession full portrait of a woman in her early thirties curly brunette with bangs hooded big green eyes square face weak jaw big nose full lips sly smile high strong cheekbones pale olive skin flowers floral Spanish shawl garden"
"Impressionist Spanish Modernist chromaticism Spanish decorative style nocturnal interior subject Vienna Secession full portrait of a woman in her early thirties medium long curly brunette with bangs hooded big green eyes square face weak jaw big nose full lips sly smile high strong cheekbones pale olive skin flowers floral Spanish shawl garden"
When I tried to get the generator to add my nose ring, all it did was add gold elements to the background.
"Impressionist Spanish Modernist chromaticism Spanish decorative style nocturnal interior subject Vienna Secession full portrait of a woman in her early thirties curly brunette with bangs hooded big green eyes square face weak jaw big nose with pierced nose and gold nose ring full lips sly smile high strong cheekbones pale olive skin flowers floral Spanish shawl garden"
At this point I decided that I had my image from one of the earlier prompts, but wanted to try the same prompt in different generators.
Stable Diffusion 2.1 Demo
Stable Diffusion 1.5
Stable Diffusion 2.0
Stable Inpainting 2.0
Stable Inpainting 1.0
Midjourney
It was actually the free and supposedly less robust version of stable diffusion
that produced the only other image that I felt was a contender for my final image.
It was also remarkable how different the results were from generator to generator,
they each had a distinct and recognizable style, which even came down to color palettes.
This may have been influenced by the input for my eye color, but all the versions
of stable diffusion that you have to register for leaned very golden green,
although the freely available version was very colorful. You can also see in a few
images a fake signature that looks almost as if the AI is trying to sign it with
Hermen Anglada Camarasa's name. Midjourney created images with lots of purple.
It’s easy to put in simple prompts, but there’s definitely an art to truly get
the image you want out of these generators. Ultimately, you’ll never be able to get
exactly what you are looking for, though that can be a fun part of the process if
you are open to discovery. In the end, this is the image I settled on.