Friday, September 22, 2023

AI Fun: Fantastic Four

One of the things I enjoy about using AI to generate images is the challenge of pushing it to its limits. Just 1 year ago, the images made from AI looked like images created by a computer. A terrible, low-memory computer. We're talkin' slightly better than a Commodore 64. But that's all changed now. Those limitations are fading fast. Many odd, weird, and disturbing images are still being made, but that is usually not the technology's fault. We have to get better at communicating with the AI technology.

Playing around, I decided to see what I could create using imagination, known characters, and pushing the reality level as far as I could. My subject: The Fantastic Four.

I started with Sue Storm. Why? She's the heart, commited, powerful, and a very sexy female hero, which, let's face it -- creating sultry women is generally 80% of what image generators are used for. It only made sense that Sue Storm would be the easiest of the four since she was probably the most normal-looking "person" of the group - I'll explain as each character is presented.

In the classic plastic/aura/ghost style often seen in comics/animation, Sue Storm was a challenge to make invisible. So I opted to show her more impressive power: forces fields.
It could be better. I would have liked all force fields to be circular. I also took the opportunity to hide her poorly generated left hand with another forcefield bubble. The explosions inside each forcefield bubble had to be generated separately. And. of course, the Circle 4 logo was dropped in Photoshop to give it proper lighting. (Assuming everyone knows type and logos are still a nightmare in AI as of this writing.)
Overall, I loved the dynamic power in this image. The suit, though unique, still fits the current trends of astro/space modern outerwear the characters usually wear today.

Next was the Thing. No doubt he is the fan favorite of the group. But what a task to generate an orange rock man with super strength.
First, it could not understand the rocky pattern of Ben's skin. Often, it looked like waffled crosshatch lines. I finally got to the closest I could with a very dynamic, powerful background:
In this image was the power and rage I wanted. There is so much going on in this image. AI often uses the colors/materials written in a prompt... EVERYWHERE. Mention "golden," and your entire image usually has a golden hue or gold items in the background. But in this case, the stone behind The Thing seems similar to his skin. It took his skin description and added it to the environment. But I like the overall look and feel.
One disappointment, I couldn't create his iconic brow. AI fights against making items it considers not "normal" (aka "popular" or "beautiful.")

The next challenge was The Human Torch, a man on fire, Johny Storm. He had to be engulfed in flames yet seen as a healthy, powerful young man. He also had to be flying. Most test images had the person on fire screaming because they were on fire.
Getting a burning man to properly float in the air is harder than you'd think. If you prompt Superman, AI usually knows how to place the Man of Steel in the air. Again, test images usually made Johnny look like he was falling -- and the lack of a fire trail didn't help sell the flight movement. So, in this image, I opted to have him float in place, leaving devastatingly scorched areas all around him.

Finally, the boss and the hardest one to create via AI: Mr. Fantastic Reed Richards. The challenge here was, again, forcing the AI to do something it is often dictated not to do: stretch and elongate body parts. Anyone who has tried generating images knows you push the prompt too much or confuse your meaning -- you get really deformed bodies, duped clone heads, and bizarre extremities. This has been corrected so often that now when I want deformed and oblong details, it is not easy to achieve. (At least at the time of this writing.)
I had to coax each area of his body to stretch, but it would only do this for a small portion of the image at a time. So, having the left-hand fingers stretch off the frame made sense. I thinned out the legs, making them appear longer than usual. The right hand is distorted, reaching for a power seed of the ultimate nullifier. I had intended on his left-hand fingers stretching over his head in the distance, but that wasn't in the cards with ol' AI.

And that-- that was my attempt at doing AI Fantastic Four.
Quite the challenge, as you can see several other users' examples below attempting to do the famous team. It's funny how so many of these images have The Thing on fire, two females, or one missing member. The last image is my favorite. It looks like a low-budget BBC sitcom from the '70s. Co-starring a "Jenny" Storm and what looks to be an overly-mutated Teenage Mutant Ninja Turtle.
I doubt if I could generate two of these characters into one image -- never mind having all four together. Another challenge for another day.

Thursday, September 14, 2023

Website Review: Ideogram.ai

I recently found this free AI generator called Ideogram.ai. What makes Ideogram.ai unique? It handles text better than any other AI generator out there right now. It is still in its infant stage but creates images with the words you prompt (well, mainly.) Check the samples below:
As seen in the samples, it can create phrases, alter font styles, and mimic styles you refer to in your prompt. It can even design the type to look like the item you are describing:
Now, its flaws are usually related to space and image size. Use large words that can't fit properly into the area you requested, and you might see letters missing from that word. Like anything with AI, practicing to see what is understood and what causes errors will help you achieve the visual goals you set in your prompts.

Currently, the website is free. There is a timer between prompting that stops people from abusing the system with multi-requests -- so you do have to wait sometimes to submit your prompts. Like most AI generators based on Stable Difusion, you get 4 sample images per prompt. You can select the images you like and download them or share them with others.

I highly recommend you give Ideogram.ai a try. Link: https://ideogram.ai/