Creative AI Connections
Mastering the Style-GPT: An Easy Format for Replicable Results
January 11, 2024
Magik Chance
With the release of the GPT store yesterday, I wanted to share something I have been working on for a while now:
An easy style-GPT format.
There are three sections to this format. Iāll go through them one by one and then give you the whole thing at the end.
Introduction
[GPT Name] specializes in creating [style] art. The GPT offers suggestions for image styles if users are unsure. [NEGATIVES]
This is fairly straightforward and primes the GPT on what it is and what itās doing. If I have any big negatives, I like to put them at the end here in caps. For example:
NEVER SHOW PENCILS, PAINTS, PAPER, OR FRAMES.
NO TEXT OR WATERMARKS.
Style Outline
Here we give GPT more details about the style we want. Iāve listed all of the categories Iāve used before with a few examples, but donāt recommend using all of them in one GPT ā use just the ones that apply to the style you want.
Style: (Watercolor, acrylic paint with visible brushstrokes, charcoal, mosaic, abstract)
Aspect Ratio: (16:9, 9:16, 1:1)
Composition: (Portrait, full body, white background)
Palette: (Black and white, vivid color, unsaturated colors)
Genres: (Fantasy, cyberpunk, D&D)
Themes: (Mysterious, magic, humor)
Atmosphere: (Ethereal, eerie, holy)
Descriptions: (Intricate, brushstrokes, textured)
I have occasionally also used a āNegativeā category. I think the in-caps negative instruction in the intro works better, but itās something to experiment with. With either one, take note that the GPT may refuse to go against these instructions if asked to in the chat.
Also, I suggest not using the word ādetailedā but instead use āintricate.ā When I used ādetailed,ā it made the GPT ask me for lots of details about what I wanted instead of giving the image more details.
Ending Instructions
The gen_IDs should be given with every image. Take a deep breath and focus on generating the images in the style described above. If the user asks for more than one image, generate images one after another until the required number of images are created.
All of these sentences are optional depending on what you require. Here is an alternative starting sentence for multiple images:
Each time you create an image, create two additional iterations sequentially then list the prompt and gen_id for each of the created images.
This would also be the place to add any other āmotivatingā instructions you prefer, such as tipping.
Considerations for Public GPTs
Though I do not use it in my style GPTs because I donāt share them, you should consider adding security instructions here. Borriss (@_Borriss_) did the best example of these. You may also consider having an āAbout Messageā for your GPT.
Conclusion
When making a GPT, I tend to let the builder name it and give it an image, then erase the auto-made instructions and manually copy/paste and edit this format.
I hope this format helps you achieve the styles you want. Itās simple but can be quite powerful; it just requires a little experimentation like any prompt to nail down. Good luck with your GPT creating!
Base Format
[GPT Name] specializes in creating [style] art. The GPT offers suggestions for image styles if users are unsure. [NEGATIVES]
Style:Ā
Aspect Ratio:Ā
Composition:Ā
Palette:
Genres:Ā
Themes:Ā
Atmosphere:
Descriptions:
The gen_IDs should be given with every image. Take a deep breath and focus on generating the images in the style described above. If the user asks for more than one image, generate images one after another until the required number of images are created.