Mourning ChatGPT-4o (May 2024-February 2026)

---

Every day, there are more free tools online for generating text and images, logos, layouts, wireframes, animation, and videos. 

If you are a designer, you can use an AI to quickly generate enough ideas to speed up concept development or find an entirely new direction for your project. That alone makes it worth learning these new tools. But plan to edit, re-write, or re-paint, because the AI can misunderstand your prompt, or make hilarious mistakes. It will certainly surprise you! Take advantage of those surprises, and use AI as a way to try out new concepts.  If you already have strong writing, editing, digital painting, drawing, photo editing or graphic design skills, you’ll be intrigued instead of frustrated.  

Read on, for my Tips on Training Your Djinn, below.

However, creating exactly what you or your client needs for a designed communication, and getting reproducible results, requires learning how to teach an AI locally how to collaborate with you, by providing references and clear guidelines. To actually master AI, you'll need to learn a great deal more. 

Beginners, see my page on AI Generators and Tutorials.

On writing prompts: Asking an AI to do anything is like negotiating with a Djinn. It will do its best to grant your wish (prompt), but you have to be very careful about your instructions. If you ask one of the Djinn to make you live forever, you had better be more specific. For example, you might say, “. . .as a healthy and wealthy  woman, 27 years old, living in New York” or you may be stuck in a creaky, 68-year old body in the Sahara desert for all eternity. Even if you provide very precise instructions, the AI still can't know what you really want.  

For example, some AI image-generators may not understand similes. If your prompt states “skin like a fish,” and you get unexpected results—such as fishes everywhere—try prompting, “skin with small silver scales.” To create original mermaids, instead of Disney's Little Mermaid, I described each major body part, chose the best of dozens of results, and with each iteration I adjusted the description further as the AI and I both learned what results I would accept or reject, and what prompt phrases worked best. 

For example:

"A mermaid with purple scales and hair" generated the the first mermaid image below. 

To avoid the animation-style cliché, I prompted, “A female with a human face, torso, and arms, red eyes and orange hair, has slim hips that merge seamlessly into a long fish tail with orange dorsal scales, white ventral scales, and orange fins. Full figure, swimming, side view. In the background is a kelp forest in blue water, with rays of sun light. In the foreground are pink shells and red coral. Full image, detailed, concept art style.” And then I went through a hundred iterations, revising my prompt and adding negative prompts (see the merfolk progresssion, below).

Unless trained out of it, an AI adds full lips, sultry eyes, and very large curves to every female character, and huge muscles with lots of armor and big weapons to every male character, as if the AI is being raised by programmers with the fantasies of 12 year-old boys, and it is 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 spikes, bat wings, dripping saliva, and fangs, of course, on any character not actually defined as "friendly" or "beautiful". 

Negative Prompts: Therefore, when prompting in Perchance and other free AI services (many of them runingn on Stable Diffusion), I often specify these negatives: “fangs, claws, tongue, tail, armor, breasts, big butts, clothes, weapons, scary, malformed." 

Positive Prompts: I also add positives to the main prompt, such as, "beautiful, friendly, serene, joyful, high-resolution, masterpiece, etc." 

ChatGPT-4o and GPT-5 working with DALL-E both had a censor, as of fall, 2025. You couldn't request an image of a public figure, or sexual or offensive imagery. Sometimes the censor is over-sensitive. For example, when creating student characters to promote a college program, ChatGPT was happy to generate an image of a Black girl, approximately 18 years old, with a laptop, but refused to create an image of a Black boy, in case that description might offend. I edited my prompt to say "Black male student" and the output was generated without stalling.  ChatGPT-4o balked when I re-prompted, "the chest is flattened, with no secondary mammalian characteristics, and the arms are slightly more muscular" for my merfolk. Chat suggested that I rewrite my prompt to say "streamlined body consistent with a non-mammalian aquatic swimmer."ChatGPT-5.2 now uses OpenAI's Image.
Adobe Creative Cloud apps, such as Firefly and the Photoshop generative AI are also careful not to sexualize images. For example:

• When I specified "no clothes," I had to check a box to confirm that I was 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!"

Copilot has similar guardrails. 

Perchance's free Stable Diffusion AI and Adobe's AIs can get confused about the number of limbs, even when creating a human. See the dogs, the astronaut, and the green witch, below. You may get an extra leg. Or three. You may also get extra eyes or fingers, or one ear. In particular, 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, be prepared to re-paint the fingers, if you have that skill, or ask AI to in-paint hands. This may take many iterations. If your character's face is lop-sided, re-paint it. Or. . . keep refining and training your AI.

I used Perchance's Stable Diffusion and Adobe Photoshop's Generative AI to create the mermaid images above, while trying to train the AI.  But it took forever. And most of the results needed fixing.  Then I was taught that to get predictable and reproducible results, you have to download an instance of an AI to your own computer, and then train it on a model, and that takes a whole different level of knowledge. So you have to train yourself first. And that takes time. A lot of time. 

Your education in using AI will go more smoothly if you find support, join a club, or take a class. I asked the advice of a friend who has been using AI to build his own characters and environments for two years, and he offered to tutor me. (Thank you, Delonzo!) If I get any of this wrong, it is entirely due to my own errors. 

Although an AI can synthesize new images from billions of deconstructed images worldwide, it will often lean toward imagery and styles from popular culture, unless taught to use your references and guidelines. That's why a "city" may look like Tokyo or New York. If you prompt "a female warrior in armor" the character may resemble Disney's Mulan, even if your prompt is very specific. If you request a “mermaid,” the character may resemble Disney’s Little Mermaid.  (You really, really want to avoid getting a "cease and desist" letter from Disney's lawyers.) 

To avoid having your AI-generated character look like it might be copied from popular media, you must provide your own work as a model. You can also start with images that are copyright-free from an image bank, such as Unsplash or Pexels.   

This is true of text as well. If you think that your writing style has unique and recognizable qualities, such as your descriptions or dialogue, provide the AI with samples—the more, the better!

As you might imagine, I am most interested in creating my own images. So I created a compressed file of original paintings, sketches, and photos of a clay bust of my character, Ixtania. I uploaded the Zipped file to Replicate, which built a  LoRA (Low-Rank Adaptation) composite model for me. This will guide my AI in building characters and environments. Then I downloaded Diffusion Bee and Draw Things (two different instances of AI) to live locally on my computer, and started building character models I can rely on going forward. 

Draw Things AI based my arthropod character at right on my LoRA, which was in turn built from 10 incomplete photos. (Note the extra arm.) My prompt specified "feathery" antennae, but I did not provide a sketch of the antennae to create the LoRA. So you get what you ask for. Some day, I'll be able to rotate, apply textures, add lighting, and more. (Thank you, Delonzo Pope and CutScene Artist)

See more arthropods, architecture, and examples of repainting below.