Some tips on styling in creating images using ChatGPT

How do you get the style right when prompting for images in ChatGPT

One thing is that you from nothing but a text prompt can make LLMs like ChatGPT/Dall-e 3 create stunningly looking images, but often it is sort of out of control.

Merging the great text understanding of a Large Language Model like ChatGPT with an image generator like Dall-e 3 is a genius move IMHO - you can "talk your way" into the image you want.

I want to share some great patterns I discovered in my journey to becoming a super user of generative AI. Here I will share there.

Get the generative id (seed) of your generated images

Sometimes you get an image which has a great style to it, something which you like and want to reuse.

What I do is to ask for the generation id: Please give the generation id for this image, and you will get a unique number like: 12mRB34VZjntNwiQ. I then save the generated image together with this id in a document in Apple note app. This then becomes a repository of styles with examples (you know: an image says more than 1000 words)

Enter a dialogue with ChatGPT to get the perfect prompt

Say you would like to create something in the style of a Rembrandt painting. Here is an actual dialogue I had with ChatGPT:

Me



So summarize in 100 words the style, so I can use it for a
prompt to generate an image

ChatGPT



Rembrandt's painting style is distinguished by its emotional depth and focus on realism. His technique utilizes chiaroscuro, with dramatic contrasts between light and dark, to create volume and texture. His brushwork is expressive and often features impasto, which adds a tactile quality to surfaces. Rembrandt excelled in capturing human emotions, giving his subjects a sense of dignity. His compositions are complex, with thoughtful use of lighting to guide the viewer’s eye. His landscapes are atmospheric, with subtle light effects, while his portraits are intimate, deeply insightful, and introspective.

I then had a dialogue where I ended up with the prompt for generating a woman eating a pizza in the kitchen "Rembrandt style" The prompt became:



Generate an image: A woman eating a pizza in a kitchen,
portrayed with the emotional depth and realism characteristic
of Rembrandt’s style. The image is infused with chiaroscuro,
enhancing the three-dimensionality through stark contrasts of
light and shadow. The painting captures the intricate textures
of the scene, from the woman's expression to the pizza’s
cheesy surface, using expressive brushwork and impasto
techniques. The atmosphere is intimate and introspective,
focusing on the woman’s interaction with her meal, and the
composition is thoughtfully arranged with subtle lighting that
guides the viewer’s eye.

Which however gave an image with the woman eating a slice, yet no slice was missing from the pizza...

And here comes a reminder: you can "prompt on" and asking for correction to already generated image:

Me



Nice, but she has taken a slice, while the pizza is not
missing a slice! Please correct that error

Which then gave me an image matching my desired image:

I then (of cause) asked for a generation id: 12mRB34VZjntNwiQ

Uploade the image you like and make ChatGPT describe what it sees

Another path to get the style you want is to upload a picture of the style you desire. You should then ask something like this:



I really like the style of this image - can you give a
detailed prompt of the STYLE and method used in this image, so
that I can prompt a generative AI and get similar style.

An example

I found this image generated using Leonardo.AI which is kind of inspiring to me.

ChatGPT then replied:



Create an image in a hyper-detailed illustrative style that
merges elements of realism with surrealism. The artwork should
have two halves with one side depicting elements of nature,
wildlife, and organic textures, and the other side showcasing
urban elements, architecture, and geometric patterns. Both
sides should seamlessly transition into each other, forming a
cohesive whole that juxtaposes the natural world with human
civilization. The colors should be vivid and rich, with a
depth that suggests a three-dimensional effect, and the lines
should be crisp and refined. The overall feeling should be one
of a thought-provoking dichotomy, blending the living with the
inanimate in a single, harmonious portrait.

Now let's try to generate an image which combines the desire to get an image of a woman in the kitchen eating a pizza in that style!

Me



I need a picture of a woman eating a pizza, Create an image in
a hyper-detailed illustrative style that merges elements of
realism with surrealism. The artwork should have two halves
with one side depicting elements of nature, wildlife, and
organic textures, and the other side showcasing urban
elements, architecture, and geometric patterns. Both sides
should seamlessly transition into each other, forming a
cohesive whole that juxtaposes the natural world with human
civilization. The colors should be vivid and rich, with a
depth that suggests a three-dimensional effect, and the lines
should be crisp and refined. The overall feeling should be one
of a thought-provoking dichotomy, blending the living with the
inanimate in a single, harmonious portrait.

(Generation ID: tWwVb30ug4EZwlPj)

Summary (done by ChatGPT ;-))

As we wrap up this creative exploration, remember that the key to unlocking AI's potential lies in our prompts. Our journey with ChatGPT and Dall-e 3 has shown us that with a sprinkle of imagination and a dash of dialogue, we can shape the raw clay of AI's capabilities into a masterpiece. Hold onto those generation IDs like treasured maps to past styles. Engage in the back-and-forth banter with AI to perfect your vision. And when you stumble upon a style that captivates you, upload it, dissect it, and learn from it. Each interaction is a step toward mastering the art of prompt engineering, transforming you from a mere spectator to a maestro of the digital canvas. So go ahead, prompt with purpose, and let your creativity soar to new heights.

Blog