For Sarah in the Secret Garden I time-boxed the AI generation part to one hour. That time was for both the text generation and the ai image generation that went along with it. That gave me very little time, but turns out to have been enough. Afterwards, I went back and did some minor editing and formatting for the post afterwards.
Start with Story
I started with the story, so that I could know what I needed for the pictures. The first step was getting an idea. What better way to get ideas than to ask for 10 from the GPT3.
Give me 10 ideas for a children’s story
1. The story of a little girl who discovers a hidden garden
2. A tale of two best friends – a dog and cat – who go on adventures together
3. The story of an unlikely friendship between a dragon and knight
4. A fun-filled, rhyming adventure about pirates searching for buried treasure
5. An action-packed story about street hockey players fighting for their neighbourhood’s coveted championship title
6..A heartwarming story about homeless animals finding forever homes 7..The rollicking escapades of three troublemaking bears at camp 8..The coming-of-age saga of twins navigating life, love, and wacky family antics 9..Mystery surrounding the disappearance of jungle animals 10…Diary entries from various woodland creatures during one very eventful day
With many lists like this the AI will break down partway through. From generating lots of test lists, I’ve found that the longer you ask it to go in a single submission the more garbled it will become. If you were to increase this to asking for 20 ideas, it would have gotten even worse. If you fix the formatting for 6 to 10 then ask for 20 it will usually work just fine.
AI Image Generation
Now we had a story idea so I started a Runpod generating the secret garden. For the first image I asked for a garden, and generated 50 images at 20 steps. From those, I chose the one I liked the most from that set and didn’t change much about it.
The story gave me a garden with fairies, flowers, birds, and bunnies. Generating four concepts is hard, and birds have a tendency to have lots of extra bits or unformed parts. I was able to edit an image to get a nice fairy and bunny surrounded by flowers. The fairy is larger than I’d like, but I’ll just say it is closer to the camera, but actually quite small. Afterwards, I took all the image steps and compiled them into a gif. Let me know if you’d like to see more of these or ideas to make it easier to see what is being edited.
More Story
I continued with the story, trying to make the old woman into the villain. The AI flatly refused to make her evil, and I was running low on time for the story. I still had a goal of two more images to get through! She now just got a little upset at not being able to use the magic of the fairy.
Back to the AI Image Generation
I decided the old woman and the house of windows would be a good subject, but getting those two together was going to be hard. I started with just a person next to a house made of windows. Then, I edited the person to fit the description of the old woman a bit better. I don’t love this image, but it works for what I wanted.
I had one more image to make, and this one I was pretty sure I could get right the first time by just generating a few. The subjects were singular enough that it wouldn’t require combining concepts. Sure enough I got a decent one on my second set of ai image generations.
Overall I enjoyed having the time restriction as it led to great concentration. However, the content did suffer compared to Little Frog, Big Dragon with the images being rougher and the story not having as solid of a plot.
Related Articles
How to use GPT-3 To Write Original Stories
Using in-painting in Stable Diffusion to make images for Once Drop in the Ocean.
Using img2img in Stable Diffusion to make the header art for Little Frog, Big Dragon.