Daz AI Studio: Prompt Engineering Guide
The following is a guide on how to steer Daz AI Studio in the right direction. Here we will go through some basic concepts of prompting then explore a few more advanced options. Hopefully, by the end of this guide, you will be able to build upon this information to start taking more control over what the generator is producing.
Prompting
A prompt in image generation is simply a word or string of words that the generator can associate with concepts to produce an output generally in line with the described input.
A prompt can be simple or complex but the effects of more complex prompting might not always be as clear as one might think.
The simplest form of a prompt would be a single word that defines a concept. Things like:
The following images were generated with nothing but the single prompt word on stable diffusion’s XL model.
‘Woman’
‘Pants’
‘Dog’
‘Castle’
These single word concepts can be used to generate a pretty loose image with the defined concept however in practice this will rarely result in matching the users expectations, normally we have an image in our head that we are trying to reproduce and we need to provide more information to the generator to start to hone in on the real picture we are trying to paint.
What’s Built-in
In Daz AI Studio, we have done a lot of this work behind the scenes already by implementing default positive and negative prompts as well as training a fine-tuned model to act as a base on which all prompts are built. This built-in control helps maintain a consistent look at the generator’s output and attempts to fix up some of the common issues with AI generation. Furthermore, our secondary models let you inject specific concepts into the generation, these work off of fine-tuned models as well as further prompt engineering.
Consider the 4 examples from the previous section, shown here with our built-inbuilt in control.
‘Woman’
‘Pants’
‘Dog’
‘Castle’
Even with these built-inbuilt in features, there is more we can do to help us guide our generation, in the following sections we will cover more in-depth concepts around prompt engineering.
Natural language
Daz AI Studio is built around the concept of natural language prompting, that is to use simple straightforward sentences without complicated syntax engineering.
Natural language prompts are things like:
The following images were generated using all built-inbuilt in prompting and the Daz base model on stable diffusion’s XL model.‘
A woman running in the park’
‘A man driving to work’
‘A dog jumping for a flying disc’
Specificity
In image generation, the words we use matter a lot more than we might initially realize. When we communicate in the real world, we often make mistakes in our speech, we use colloquialisms that seem perfectly understandable to us but are influenced by our own biases. In reality, while your close circle or people who have known you might understand what you are referring to, if you were to say the same thing to someone you met on the street they may have no idea what you are talking about. We need to treat the image generator like a person we’ve never met. Using broadly known or acknowledged concepts is more likely to generate desired results than terms that are sequestered or understood by only certain groups.
At the same time if we have an idea in our head we want to try to be as specific as possible and avoid broad terms where they aren’t sufficient. A simple example would be using the word ‘person’ when we actually want to see a ‘man’ Another example would be using the word ‘castle’ this may generate all sorts of castles but if we are envisioning a dark sinister castle that count Dracula might inhabit in a horror film, then we would want to specify:
a ‘haunted castle’
or a ‘vampire’s castle’
or ‘an ominous gothic castle’.
Impact
Certain concepts will have more impact than others when crafting prompts for image generation. Try playing with the following concepts to add some more depth to your images.
Genre
Defining a genre is a great way to get some life out of an image. What setting or world do you want the output to exist within? Using genre concepts can achieve this. Words like: Steampunk, fantasy, science fiction, and post-apocalyptic are a good broad start.
Backgrounds
If you find yourself getting stuck with lots of images with plain flat backgrounds you can start adding in background prompts as well. Try using things like: forests/forest, city streets/street, tropical islands/island, and deserts/desert.
You can even try things like ‘spooky background’
Color
Color can be unruly in image generation, you can’t always control what the color term applies to, however, you can give a general palette direction to your image. Simple colors work but if you want more complex palettes try things like: Earthy colors, Pastels, neon, and noir.
Lighting
You can control the lighting in your images as well. Try terms like: ambient, rim lighting, lens flare, back-lit
Mood and Atmosphere
What do you want your image to feel like? You can use descriptors and adjectives to help direct the mood of your image. Trying things like: eerie, heavenly, horror, and whimsical in your prompt can direct the feeling.
Action
Actions are things that affect the subject’s motion. Verbs like running, flying, singing, dancing. These words don’t always yield the results you may expect but they can help you push away from the standard output of a person standing straight in the middle of the frame.
Camera Shots
You may have heard terms like ‘close-up’ or ‘wide shot’ before in reference to photography and videography but there is an extent to which image generation can able to parse these terms.
Panoramic shot
Example Prompts
‘a man jogging on a path, sunny, tranquil, in a park’
‘a dog running on the beach, long shot, the ocean in the background, overcast’
‘a woman dressed in steampunk attire, gloomy atmosphere, rust, and copper tones’
‘a cinematic close-up of a monster, with red eyes, dark, ominous, horror’
Advanced Prompting
The following section will try to cover some more advanced tools we have available when crafting our prompts in image generation.
You may have seen examples of this kind of prompt weighting before, with long strings of text used to achieve seemingly perfect results from highly skilled prompt engineers. While the Daz AI Studio is designed to try to simplify the process, we will try to demystify it here so you can play with these concepts in your own generation.
Tagging
Tagging is an alternative to natural language prompting that can be used to list out concepts in a range rather than by constructing sentences, this type of prompting can be helpful, especially in the case of larger more complex prompts.
In tagging prompts, we separate concepts by commas instead of connecting them with strings of contextual language. These work best when you want to see specific things in a prompt but are less concerned about how they interact. They can also save space in larger more complex prompts as there is a limit to prompting before they are truncated for generation.
‘A Puerto Rican man with a short well-groomed beard is sitting in a cafe drinking a latte in the early morning’
‘Puerto Rican man, short beard, well groomed, cafe, drinking a latte, early morning’
Features
Daz AI Studio has a few other options that are not directly related to prompting but can also help with controlling your results. We’ll outline these below.
Prompt Strictness
Your selection here will have a subtle effect on the output of your image. At its most basic explanation, the CFG scale determines how closely the generator adheres to your prompt. Choosing the highest value will push the generator to follow your words more closely and choosing the lowest setting will allow it to have more freedom.
Seed Iteration
Once you find an image that mostly matches your expectations, it can be helpful to lock in your seed and continue iterating on a generation that will only be subtly affected by the alterations in your prompt.
Multi Image Generation
This feature can help in your generation simply by having more variations to work within in each iteration.
