By gerogero
Updated: November 16, 2024
AI generators are simple enough: you type in a prompt and then click “Generate” to get your image.
The question people have when they are starting out, is not ‘how do I generate’, but ‘how do I generate better images’.
This guide aims to answer that. It is the result of constant experimentation, and learnings from the BetterWaifu website, on which hundreds of thousands of images are generated every day.
To anyone trying to improve their prompts, the first suggestion in our Discord server is usually to browse Danbooru.
But for the first-time user, Danbooru is confusing as hell. It just looks like a messy jumble of images and tags. So how does it actually help us generate better waifus?
Danbooru is the largest anime imageboard in the world. Without exception, all anime AI generators use its images for training. There are many ‘booru’ sites that use the same layout, such as Rule34 and Gelbooru, but Danbooru is the largest.
On Danbooru, volunteers tag all images comprehensively by their content. These rags range from physical characterstics “large_breasts“, “red_hair” to objects “book” to actions “fellatio“.
You can use these tags directly in your prompts. (there no difference between underscore and space in prompts)
The key idea: If a tag has a lot of images, it means there was a lot of content to train the AI on that tag. Thus, BetterWaifu and other AI generators can usually generate it.
So effective prompting is a process that starts with finding good tags and using them directly in your prompts.
Start from the big Tag Group list. Here I’ll feature some groups that are particularly interesting
Best tags to specify breast sizes, visibility, clothing for breast-focus
Long list of assorted sex acts
Sex positions including bondage positions
Types of nudity, partial nudity, swimsuits.
Background colors
Types of lighting, can greatly impact your result
Angles, perspective, composition
Facial expressions, emotions (including sexual), facial features
Don’t know how to start? Go and see the Tag Groups and think about what is most important visually in your dream image.
I usually like to follow a format by using 5 groups of words, with sub-categories inside.
I try to separate those concepts with newlines, it helps me to quickly make adjustment. The newlines do not affect the generation.
Keep in mind that it’s a very rough guideline. It’s ok to mix this order, especially inside a category.
Wait, isn’t that a big long and complicated?
Yes it is! But you don’t have to have any category you don’t want, and a single word can suffice for a category. It’s really about what level of control you want. Let’s take a look at some examples.
I’m thinking of a black maid outfit shot from below. The background doesn’t matter but I’d like cool cinematic effects. For inspiration, I would click on the links above in each of the categories and subcategories I’m interested in.
Composition: from_below, Subjects:
The prompt is split by words (or chunks of words) to transform them into numerical representation called tokens. Depending on which models are used and how tokens are normalized, certain parts of your prompt will have more or less attention. My rule of thumb: the longer the prompt, the more control you’ll have over the whole prompt.
Here’s some resources if you want to read more on that matter:
Tokens with SDXL and SD15 models – Alen Knight
Token normalization & Weight interpretation – BlenderNeko (Github)
Here is an example following the structure. I usually put some break lines to have a better look on the main areas of the prompt.
sidelighting, light particles,
1girl, ginger, solo, smirk
sweat, freckles, small breasts,
ginger hair, long hair, straight hair,
blue eyes, glowing eyes, glasses, looking at viewer,
smile, smirk, grin, frown,
white tank top,
indoor, library,
sunset, sunny, daylight,
desk, chair,
As you can see, the breaklines don’t divide perfectly the 5 fields of the structure. Depending on what you want to generate, it can make more sense to visually separate smaller and longer parts. Here, the subject, actions and posture of the body are grouped together: splitting them would create a mess visually more than anything.
Here are some usecases and favorite keywords. I try to organise by logical group, from the most general to the more specific. I recommand to cherry-pick the one that make sense for you rather than copy-paste the all line.
You’ll also find a reference to where I place it in the structure. I’ll use an ID with the format [xx.yy], using the acronym of the category and its sub-category.
Keep in mind that some of them could also go higher or lower in the apparition order (e.g. upside down could be placed as [compos.pov] or [body.posture].
Here is what I use in almost all my prompt to influence the overall quality of the image. I usually add additional negative keywords if something that I don’t like appear, but not before. Again, I like to keep things short and simple.
Following list cover some classic cases, detailing both positive and negative prompts.
Note: Negative+ is an additional list where I cherry-pick according to the situation. In my experience, negative keywords related to anatomic don’t always improve the results, so I just keep them aside in case I need it.
On BetterWaifu the “Auto-settings” option automatically chooses preset optimal settings after you select a model. These settings are de...