Image Prompt Writing

Text to image prompting is an odd beast. 

The difference between what does work and what doesn't work varies depending on how old the model/checkpoint you use, or which datasets it uses, what style of captioning it uses in the dataset and also which algorithm you are using and lots more as well.

Each prompt we write whether we are following a template consciously or not, can be divided into different sections and all prompts are made up from one or more of the following sections clumped together in one way or another. This is a process I use 

Your most basic prompt is a simple subject matter/theme. This is the focus of your image. So, we can describe this concisely here. For this example, I will use.

"A dog running"

So, as you can see my prompt has a subject matter and a theme in a few words. 

The secondary subject matter can be used to describe the background, other characters, or the atmosphere. We can introduce more detail to our prompt. So, I will add to my prompt.

"Splashing in puddles, on a rainy day"

Now we have added more details to our main subject. We have atmosphere and weather effects. We have added a dynamic element by adding puddles and rain.

This section is especially important and can be divided into three subsections which are:

With artists: so, your prompt would read something like "By Kermit the Frog”

Without artists: so, you might describe two style types "grunge style rough brush strokes with biomechanical carbon fiber aesthetics."

Or, artistic movements: so, you might just use something like "portrait impressionism, sci-fi horror."

We can use all three of these in one prompt or just two or just one or any combination of them. Each will work differently with different models. 

Here we can get technical and include specific tools and techniques used in the creation of images. Things like computer programs and other quite specific details like camera details etc. so for example we could put them in:

"Shot on Nikon, DSLR"

Add the extra details that enhance other parts of the prompt. 

Here I usually like to use up to ten keywords which you can use to enhance the details of any of the previous parts. In many ways I use this section to reinforce parts of the prompt and to add depth and detail. It can be part of a poem, a song, words that are related to subject matters or theme or style. So, for my example prompt I am going to use:

"Stunning action, splashing puddles, joyous life-like candid details, breathtaking intricate reflections"

 So, as we can see from my completely blind example, we can create a good prompt by following a structure. You can use this workflow for any model. Different models work better with different combinations of these steps. Remember you can use just one step. Or just two. Any two. See what you get. Think about which of the other steps can enhance what you are trying to make.    

So, we have looked at the distinct parts of a prompt. Here is the cool thing. You can use just any one part of the process above, or use all of them, change the order you use them in for different effects and shift emphasis on distinct aspects.

It is a process I have been using quite successfully for over a year now. Every single prompt is made up from one or more of these types of sections. 

You can write simple prompts using just one step of the process. 

The image shown is the image made in Flux on the NightCafe webpage.

Do you need to use a negative prompt to refine the image? 

Do you need to change artists etc., we can change the order of the steps to shift emphasis on distinct aspects. 

Lots of other small steps to be considered as you refine your prompt. Hopefully following this process will help you to write better prompts so you are able to better visualize what you were trying to create.