Tips to create re-usable prompts that produce consistent images
Designing AI image prompts with tuner and style reference
In this article we are going to look at how to use AI image generations tools to help you produce consistent and predictable images, and turn your hobby into possibly a business of buying and selling prompts (check the end of the article).
Why do you want the consistency?
Generating AI images for fun is one thing, but if you really want to get the best value for your money, you want these tools to be able to generate highly consistent and predictable outputs because every image you generate on these tools costs you money.
Secondly, if you are doing this as a serious profession, you want to be able to re-create same effects without having to spend too much time experimenting to speed up your workflow.
Lastly, another good reason to be able to generate these consistent and predictable outputs is the ability to sell the prompts for those who are trying to turn AI art into a profession (yes, that is a thing 🙂).
Crafting the prompt
A big part of designing re-usable prompts if to first “craft” a prompt that you really like and delivered the expected result. This is the part where your individual taste and creativity comes into play. Today, as a prompt designer , you need to experiment with multiple variations of your prompt until you are satisfied with the result you are getting. This is both time consuming and costly. So let’s talk about how we can simplify this.
A few tricks come in handy at this stage:
1. Do not over engineer the prompt
AI tools do not literally understand everything you tell it. For example, if you try to describe the ambience of a scene as: “The overall atmosphere is one of relaxed productivity and a cozy, personalized workspace” they are going to only understand and pick up keywords that they decide to chose and generate an image using them. So being precise and selective in choosing your words can help you get closer to that you want to create.
Here is an example:
Prompt (detailed):
A boy walking down the street in a happy mood while birds in the sky fly away. The atmosphere is one of joy and happiness with calm winds blowing and trees swaying in the background. Its the vibrant spring time.
Here are the results of this prompt.
Now let’s simplify this prompt a bit and see what happens.
Prompt (simplified):
A happy boy walking down the street, birds in the sky, vibrant spring time
This is what I got.
As you can see, the results were pretty impressive with the simplified prompt. The boy looks happier, the image shows more vibrant spring colors. Not sure why, but this could be a result of the LLM having to chose from fewer options to iterate.
You can get pretty close to what you want by choosing to focus on whats important for you in the image. I will not go into some other techniques such as how to specify weight for certain words or exclude certain things from the image etc. There are a ton of posts on Medium covering these topics. Here are some of the posts you might find interesting on these topics — 1, 2, 3
2. Tune prompt with version 5.2
Once you are getting closer to achieving the results where you are able to get the key elements right, you can try to further expand your prompt by “tuning” your prompt. Again, I will not go into a lot of details here as there are many articles that teach you how to tune a prompt — 1, 2, 3
So, lets take our simplified prompt and tune it to generate a specific style. In Discord, you can type /tune and press Enter followed by the prompt you want to tune.
After you decide how many style directions you want to generate (16, 32, 64, 128), you can run the tuner and it will give you the URL where you can now generate styles.
In my case I went with 16 because 3¹⁶ is about 43 million combinations :)Here’s what my result looked like. As you can see the tuner generates a lot of varying styles that you can mix and match.
I liked this illustrative style on the right hand side of option-4 and I wanted to give that a try. So I chose only this option to generate a style code ––style 2ZHUMDAL0
Next, I went to Midjourney and tried to generate some images with this style. I used the exact same prompt to generate some more images of the boy walking down the street:
The style did live up to its expectations and I got some cool variations of the designs. Next I tried to modify the prompt a little.
The results were still consistent with the style:
A young women climbing up the strairs, paintings on the wall, bright summer time — style 2ZHUMDAL0
Couple shopping at a flea market, crowded, clouds in the sky — style 2ZHUMDAL0
As you can see, the style does work. But one thing I have noticed is that if you try to apply the style to a subject far away from the original prompt on which you tuned it, you may get unexpected results like this:
A still life photography of some old utensils and some fruits with a cloth hanging on the table — style 2ZHUMDAL0
So there are certain limitations of how style tuner works. But most of the time you can achieve consistently predictable images.
3. Style reference
This is where the third technique comes into play “––sref” — The Style Reference.
Style reference works with version 6.0 and it takes an image as an input and generates the new image in the style of the input image. Once again there are many tutorials on how to master style reference, so I wont go into the details but leave you with some rabbit holes :) — 1,2,3
I wanted to see if using style reference can I still generate the still life photograph that I was trying to generate in the illustration style I had tuned.
To do that, I selected this image to be my input for style reference:
Now, my prompt looked like this:
A still life photography of some old utensils and some fruits with a cloth hanging on the table — sref https://s.mj.run/IKCt2i12w-g — v 6.0
This time I was surprised to have achieved the positive result. The image did capture the essense of the illustrative style and generated the still life image I asked it to.
You should test your prompt on multiple subjects to be confident that it will generate consistent results. If you need to tweak things, you can always go back a few steps or try new things as this is all about experimentation :)
In the end, I would also like to introduce you to the concept of selling your prompts to make money. A lot of AI image artists are realizing that there is a big void that needs to be filled in terms of artists who can craft new and unique image styles through such experimentation.
If you are someone who wants to try things out, I am building a Prompt Marketplace app called PROMPRMAKR where AI image artists can buy and sell their prompts. It’s free and available for Android and iOS devices. Currently it allows you to post prompts that others can download for free, but very soon we are going to introduce payment system that will allow you to start charging for your work.
I would love for you to try it and give some feedback if possible :)
