Compositing photos with generative AI

---

Training an AI is like negotiating with a Djinn. It will do its best to give you what you request in a text prompt, but you have to provide very precise instructions, and it still can't know what you really want. It will certainly surprise you. You have to train it, and that takes time. A lot of time. 

I've used Stable Diffusion's AI and Adobe Photoshop's Generative AI plug-in to generate these images, while trying to train the AI.  Although an AI can refer to billions of images worldwide, it leans toward imagery and styles from  popular culture, unless trained out of it. Towns usually look like either Tokyo or New York, with skyscrapers. If you request "a female warrior" you'll get characters that resemble Disney's Mulan, unless your prompt is very specific. If you request a “mermaid,” you may get characters that strongly resemble Disney’s Little Mermaid.  You really want to avoid getting a "cease and desist" letter from Disney's lawyers. If possible, provide a reference image, but it must be your own image, or one that is copyright free from an image bank, such as Unsplash or Pexels.

TIP #1. An image-generating AI may not understand similes. In your prompt, you can’t request “skin like a fish.” You should write, “skin with small orange scales.”  To create original mermaids, I had to describe each major body part, choose the best of dozens of results, and with each iteration, adjust the description further, as the AI learned. 

For example:

“Female human face, orange hair, red eyes; human torso with slim hips merging with fish tail, orange scales on skin, orange fins. Full figure, swimming, side view. In the background is a kelp forest in blue water, with rays of light. In the foreground are pink shells and red coral. Full image.” 

And then you need to go through a hundred iterations (see Mer-people, below).

TIP #2.  Unless trained out of it, AI adds full lips, sultry eyes, and large curves to every female character, and huge muscles with lots of armor and big weapons to every male character. Perhaps it is being taught by programmers with the fantasies of 12 year-old boys, or heavily influenced by popular culture. With non-human characters, the AI can become very confused. I have seen eye shadow and lipstick on monsters, bugs and reptiles. (As well as claws, spikes, bat wings, dripping saliva, and fangs, of course.) 

Therefore, in Stable Diffusion I must also specify these negatives: 

“No legs, no teeth, no fangs, no claws, no tongue, no armor, no breasts, no big butts, no clothes, no weapons, beautiful, not scary.“ 

And yet, Photoshop's AI was more cautious about sexualizing images. If I specified "no clothes," I must check a box to confirm that I'm over 18. When I requested a golden "breast plate" on a female warrior, I got an error message and had to change it to "chest plate!"

TIP #3. Stable Diffusion and Photoshop's AI both got confused about the number of limbs, even when creating a human. See the astronaut, below, and the arthropods. You may get an extra leg. Or three. You may also get extra eyes or fingers, or none. Photoshop's AI has a lot of trouble with hands and faces. If you want your character's hands to clasp the hilt of a sword, better be prepared to re-paint the fingers, if you have that skill. If your character's face is lop-sided, re-paint it. Or keep refining and training.

Nevertheless, it's very exciting! As you might imagine, I am most interested in creating my own images. So I'm uploading original art work to an AI, to guide it in building characters and environments. Eventually I'll be able to rotate, apply textures, add lighting, and more.  More to come.