This article will talk about building prompts in Perchance's text-to-image-plugin. If you want to know how 'Diffusions' work, see this cool website!

Licensing

The Text-to-Image model uses Stable Diffusion and by using it you agree to the Licensing and Terms and Conditions stated in this license.

Per Section III Paragraph 6, Perchance claims no rights in the Output you generate with the model. You (the user) are accountable for the Output you generate and its subsequent uses.

Here are the Use-based restrictions stated in the License (see Attachment A at the bottom of the license):

You agree not to use the Model or Derivatives of the Model:

• In any way that violates any applicable national, federal, state, local or international law or regulation;
• For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
• To generate or disseminate verifiably false information and/or content with the purpose of harming others;
• To generate or disseminate personal identifiable information that can be used to harm an individual;
• To defame, disparage or otherwise harass others;
• For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
• For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
• To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
• For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
• To provide medical advice and medical results interpretation;
• To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).

Prompt Terms

First we will talk about terms that we will be using in this article:

• Raw Prompt
  • This is the initial/main prompt that you will input to the generator. It contains the things that you want in your image.
• Tags
  • These are terms, words, or phrases, that are used to modify your Raw Prompt. These add details, fine-tune the composition, or stylize your image. There are four categories of tags:
    • Content Type - These are tags that will specify the image or content type that the AI will generate, e.g. Photograph, Picture, Painting, Sculpture, Model etc.
    • Description - These are tags that will describe or define the Raw Prompt, e.g. Adjectives or Descriptors (beautiful, young, lean, etc.).
    • Style - These are tags that take style from well known artists, art movements, or general style of the image to be generated, e.g. Van Gogh Painting, Cubism, Brutalist Architecture, etc.
    • Composition - These are tags that modify the composition i.e. camera angle, centering of subject, lighting, pose etc.
• Seed
  • This is the base 'noise' in which the image will be derived from. Seeds are useful for finding effective tags by ensuring a 'similar' image will be generated upon changing the tag.
• Guidance Scale (CFG - Classifier Free Guidance)
  • This is a scale on how much the AI will try to adhere or match the prompt. It's values are from 1 to 30 where 7 is default. ? Higher Values = Closer to Prompt (More Chaotic), ? Lower Values = Farther to Prompt (More Realistic).
• Negative Prompt
  • This is the prompt that will try to subtract or remove the items that will be in the final image. This is useful for preventing things to appear in your iamge.
• Resolution
  • This is the resolution/base size of the image that will be generated. There are currently three:
    1. 512x512 which is a square
    2. 512x768 which is a portrait/vertical rectangle
    3. 768x512 which is a landscape/horizontal rectangle.

Prompt Settings

Prompt Setting Management

Based on the text-to-image-plugin page, we can do two things with prompt settings: (1) have them all inline or (2) have them in a list.

We would recommend using the list method like so:

promptData
  prompt = painting of [character] in [place], [season]
  seed = 123
  size = 400
  style = border:4px solid blue; margin-top:20px;

Compared to:

prompt
  [character] in [place] (size:::400) (seed:::123)

Since in the long run, it will be more readable in the list form compared to the inline method.

The settings that are important are (in order of importance):

1. prompt
2. negativePrompt
3. guidanceScale
4. resolution
5. seed
6. width (for portrait, height if landscape, size if square)

Here is a template of that settings.

promptSetting
  prompt =
  negativePrompt =
  guidanceScale =
  resolution =
  seed =
  width =

Prompt and Tags Management

For managing the prompt and tags, we would recommend also having them in multiple lines instead of inline or one line.

If you haven't checked out the User Input and Output Formatting article, which talks about the $output keyword, I would recommend reading through it as you will understand the following formatting easier.

The following are the benefits of having the tags like so:

1. Easier Reordering and Commenting of Tags
   • This allows the user to reorder/comment the tags fast compared to inline in which the user needs to click and drag items to delete and if the tag is effective in the prompt, they need to redo the deletion.
   • Having it in multiple lines, we can just comment the tag that we won't want and if we want it back, we can just comment it back.
   • Reordering tags in inline needs copy and pasting while in multiline, we can just use Ctrl+Shift+Up/Down to move them and Shift+Up/Down to select mutliple lines to move or comment (see Keyboard Shortcuts for more shortcuts).
2. Readability
   • Having it in multiline we can isolate tags easily compared to a paragraph of tags.

Here is an example of the multiline Prompt and Tags Management

rawPrompt = ...

prompt
  $output = [this.selectAll.joinItems(", ")]
  [rawPrompt]
  [tags]

tags
  $output = [this.selectAll.joinItems(", ")]
  tag1
  tag2
  tag3
  tag4
  ...

negPrompt
  $output = [this.selectAll.joinItems(", ")]
  tag1
  tag2
  tag3
  tag4
  ...

promptSetting
  prompt = [prompt]
  negativePrompt = [negPrompt]
  guidanceScale = 7
  resolution = 512x512
  seed = -1
  size = 400

Here notice that on the prompt, we are joining each lists with a comma and since the output upon calling the list is the joined items, it will be instantly a long list of words deliminated by commas.

Prompt Making

Seeds vs Prompt

Here is an example of difference between prompts and seeds to get variations.

On the first row, we have the same image seed but it has different prompts or tags where the color differs. However looking at it, the images are almost similar, though with some variations.

On the second row, we have same prompt but different seed. Looking at it, each image are different to each other as they have different base 'noise' to make an image from.

If you want a different image, you change the image seed but stay with the same prompt. But if you want the same image but different variation, you can change the prompt but same seed.

By the default the seed will be randomized.

Change Seed but Same Prompt = Different Image, Change Prompt but Same Seed = Different Variation.

Here is the generator used for the previous image.

Negative Prompting

Here is an example of how negative prompting affects the image.

Here we can see different seeds but with the same prompt.

Then, upon applying the negativePrompt, with the same prompt and seed, we can see that the items changed and removed or lessened the items specified in the negativePrompt while still retaining the base image essence without the negative prompt.

Here is the generator used for the previous image.

Prompt Ordering

In adding tags and negative prompts, order matters.

On the example above, we can see that if we re-order the tags/prompt that we have we can see the difference between the images.

Here is the generator used for the previous image.

For creating a good image, a usual order is:

First is the raw prompt, what we want in our image. Then is the content type, what type of image do we want. Next is the description, how we would describe our image, what details does it have, etc. Then what is the style of the image, do we want it to look like something that a popular artist makes, etc. Then lastly how is the image composed or laid out.

This tags/prompt order is just a guideline you can rearrage them on what you want to be first applied to the image.

For example, if you want to set the composition of the image first before specifying the style, you can rearrage them and see what suits your image best.

Here is a example of the prompt robot, geometric, 90s style, cinematic shot rearranged:

Here is the generator used for the previous image.

We would also recommend placing the tag categories in their own lists like so:

tag_contentType
  $output = [this.selectAll.joinItems(",")]
  ...

tag_description
  $output = [this.selectAll.joinItems(",")]
  ...

tag_style
  $output = [this.selectAll.joinItems(",")]
  ...

tag_composition
  $output = [this.selectAll.joinItems(",")]
  ...

tags
  $output = [this.selectAll.joinItems(",")]
  [tag_contentType]
  [tag_description]
  [tag_style]
  [tag_composition]

Then, you can use this Permutation Calculator to generate possible arrangements of your tag categories.

Here is a template with all of the settings to get you started with the prompting.

Guidance Scale

Guidance Scale determines how the AI will try to match the given prompt, here is an example of it.

The prompt is woman portrait, goth aesthetic, 90s style, cinematic shot. The guidance scale is from 1 to 30 and as we increase it, the AI tries to match the given prompt.

Here is a prompt: woman portrait, hippie aesthetic, 90s style, cinematic shot with steps of 5 in guidance scale (5 to 30):

Here is the generator used for the previous image.

The default guidance scale is 7.

Resolution

Resolution determines the size of the image that will be generated. Here is an example of it:

The prompt was woman portrait, hippie aesthetic, 90s style, cinematic shot. Here we can see that a 512x768 is good for portraits but 768x512 is good for a cinematic portrait shot. The 512x512 is the default.

Guidance Scale here is default (7)

Here is the goth aesthetic prompt:

Here is the robot, geometric prompt:

Here is the abstract geometric red prompt:

Here is the generator used for the previous images.

Advanced Prompt Making Techniques

Seed Hunting

Since mostly the image depends on the seed, we can 'hunt' seeds by either setting a start and end seed or generate a string of numbers to input to the seed setting of the prompt.

For example, modifying the different resolution generator to increase the seed after iterating through the different resolutions, it will look like this:

We can see that changing the seed from 100 to 104, there are significant changes on the images. Now, what if we want the image from seed: 104 and resolution:768x512. Then we change the prompt while retaining those parameters, will look like this:


We can see that the position of the subject stayed relatively the same but the overall look changed due to the prompt being changed.
Here are other examples of the same seed, but different resolution and same prompts:


We recommend that when seed hunting limit your prompt/tags to most basic/raw form. This is to have the overall feel of the 'seed' since if you have a lot of tags, the base noise of the seed may be hard to tell.

For example, you are searching for a good portrait seed, you can just have the prompt portrait and see the 'feel' of the seed with that simple prompt.

Here is the 'basic' portrait from the seed 104 and resolution 512x768 and the result of the original prompts:

We can see that the subject is already centered; which will make subject after adding other prompts also centered. It also has pretty nice brown and black colors that reflected on the images as well.

Here is the generator for the example of seed hunting (might be better to create your own)

Tags Effect Testing

Now that we have the seed and resolution, we can now test the effects of tags. For example, what would happen after a tag is added or rearranged.

The image above shows the effects of the tags after each is added to the prompt. Now what if we rearrange them:
hippie aesthetic first
90s style first
cinematic shot first
We can see that there are some subtle differences between the images. Make sure to test the order of your tags to find what you might be after.
reversed order of tags
Here on the reversed order, the main difference is the clothing of the final image, which is no longer floral or patterned. Here is using goth aesthetic:

You can also apply this to the negativePrompt, here is an example:

Here we add the painting, monochrome, cosmetics, worse quality as the negative prompt. We can see that adding the monochrome reduced the black and white on the image. Here is on the hippie aesthetic prompt:

Here is when the tag in the prompt is sequential while the negativePrompt is static and when the prompt is static while the negative prompt is not:

Here is the generator used in the previous examples.

Tags Emphasis

As previously stated, guidance scale is how much the model will try to match the prompt given. Previously, we are changing the 'global' guidance scale, but we can also change the tag's guidance scale or also called emphasis.

By default the tag's scale is 1. To increase it, we need to enclose them in parenthesis, and in doing so, it will increase its scale by 1.1, meaning 1*1.1. Upon enclosing it to two parenthesis i.e. ((tag)), it will once again increase by 1.1, meaning the value will be 1*1.1*1.1

The equation will be 1.1**N, where N is the number of parenthesis.

Here is an example of it in the goth aesthetic tag:

We can see the goth aesthetic of the image increase as we increase the parenthesis on the tag. Another syntax is the following: (tag:value), in which we can see here:

We can also de-emphasize tags, albeit sublty by changing the value to below 1 e.g. 0.9, 0.8 etc.
Here are some examples of it:

We can also emphasize the tags in the negative prompt like so:

Here we emphasize the monochrome tag and we can see that the lipstick of the portrait turned red, and the dress is somewhat purply due to the brownish color filter.

Here are both in the hippie aesthetic:

We can see that the hippie aesthetic resulted in more braids and colors, while in the monochrome, the hair turned more vibrant and the overall colors popped more.

Here is for both man portrait:

Here is the generator used for the previous images.

We recommend using the (tag:value) method since using multiple parenthesis doesn't look good for readability. You can achieve the same values as the multiple parenthesis using the equation:

(tag:[(1.1**value).toFixed(2)])

You can also group multiple tags like: (tag1, tag2, tag3:value) to have them all have the same emphasis.

Blending Tags

At Automatic1111, they use [from:to:value] to blend tags, unfortunately, we cannot do that in the model we have in Perchance.

We can blend tags using [from:to:value] by adding .map(a => a.getRawListText) in the $output like so:

tags
  $output = [this.selectAll.map(a => a.getRawListText).joinItems(",")]
  ...

To format the blended tag, we need to escape the [] in Perchance like so:

tags
  $output = [this.selectAll.map(a => a.getRawListText).joinItems(",")]
  \[from:to:amount\]

The value determines how much of the from is present in the blend. For example:

• [from:to:0] - which means 0 of the from tag, and 1 of the to tag.
• [from:to:0.25] - which means 0.25 of the from tag, and 0.75 of the to tag.
• [from:to:0.75] - which means 0.75 of the from tag, and 0.25 of the to tag.
• [from:to:1] - which means 1 of the from tag, and 0 of the to tag.

Here is an example with [hippie:goth:value] where value is from 0.00 to 1.00 with 0.1 increment:
Unfortunately, the only way to see the [from:to:value] is to hover on the image, since Perchance evaluates it upon displaying on the page.

Here is with [goth:hippie:value]:

Here is with 0.00 to 1.00 with 0.05 increment: goth:hippie

hippie:goth

Here is the generator used in the images above.

Another way to blend them is like the following:

(tag1:1) (tag2:1.2)

You can emphasize the tag that you would want to apply on the first tag. Still order matters on which tag is first and which is the second. Here is an example of it:

Above, goth aesthetic is first then the hippie aesthetic is second. Here is hippie aesthetic first, and goth aesthetic second.

Here is the generator used in the images above.

This also works for the negative prompts.

Here is another example of blending with de-emphasizing.

The images above blends between man portrait and woman portrait with de-emphasizing on the other.

Here is the generator used in the images above.

ANDing and ORing

Another way of 'blending' tags is using AND or OR on the tags, here is an example.

Here we are ANDing and ORing the goth aesthetic and hippie aesthetic, we can see that the AND tries to incorporate both while the OR tries to incorporate either. The OR method looks like the (hippie aesthetic:1.0) (goth aesthetic:1.0) method like above.

Tag Order seems to matter still.

Here is the generator used in the image above.

Here is with the images and having the tags deemphasized and emphasized (from 0.5 to 1.5) and ANDed.

Here is the images being ORed.

Here is both but with deemphasis and emphasis from 0 to 1.


using goth aesthetic first.


from 0.5 to 1.5 emphasis

Here is the generator used in the images above.

Commas in Prompts

Some say that commas in prompts to separate the tags doesn't affect the overall image. However, sometimes using commas to separate the prompt does give a significant effect ot the image:

Here we see the different effects in different comma placements in the image.

• The top right image (woman portrait goth aesthetic, 90s style, cinematic shot) seems to have more contrast on the image
• The second image from the left on the top row (woman portrait goth aesthetic 90s style, cinematic shot) seems to be the only one without a black lipstick.

Here is for the hippie aesthetic

Feel free to change up your commas and see a difference in the images.

Here is the generator used in the images above.

Add/Remove Tags during Generation

You can add and remove tags at a certain point on the generation using the following syntax:

• [to:when] - adds to to the prompt after a fixed number of steps
  • Basically means, add to when the step is at when.
  • A hill [with a monolith:0.5] means:
    • A hill - for the first half of the generation steps
    • A hill with a monolith - for the remaining generation steps
• [from::when] - removes from the prompt after a fixed number of steps
  • Basically means, remove from when the step is at when.
  • A hill [with a monolith::0.5] means:
    • A hill with a monolith - for the first half of the generation steps
    • A hill - for the remaining generation steps

The Model in Perchance is assumed to be generating at 20 steps, thus the increment would be 0.05 per step.

Here is the hippie aesthetic being added at different steps upon generation.

Here is the goth aesthetic being removed at different steps upon generation.

Here is the generator used in the images above.

Tags Per Steps

You can change when the tag is applied depending on the step. This is also known as 'Alternating Words' in Automatic1111. The syntax is:

• [tag1|tag2] - means, the generation will start with tag1, then the next step will use tag2 then it will alternate back to tag1 and so on.
• [tag1|tag2|...|tagN] - means the generation will alternate between the tags per step until tagN and it will loop back to the start.
• [tag1|] - means that generation will start with tag1, then the next step is blank, then tag1 and so on, i.e. tag is being applied every even steps every two steps: 0,2,4,...
• [|tag1] - means that generation will start with blank, then tag1, then blank, i.e. tag is being applied every odd steps every two steps: 1,3,5,...
• [tag1||] - means that tag1 will be applied every three steps starting from 0 i.e. 0,3,6,...
• [|tag1|] - means that tag1 will be applied every three steps starting from 1 i.e. 1,4,7,...

Here is goth aesthetic being applied at different intervals starting at 0:

Here is the generator used in the image above.

Here is goth being alternated with different keywords:

Here is goth being alternated with increasing keywords:

Here is the generator used in the images above.

BREAK Keyword

According to the Automatic1111, tokens are separated into chunks of 75. That is, if your prompt has over 75 tags/prompt, it would create another chunk.

We can use the BREAK keyword, to force those chunk separation without needing to fill the remaining tokens.

• woman portrait, goth aesthetic, 90s style BREAK cinematic shot would be something like:
  • ['woman','portrait',',','goth','aesthetic',',','90s','style',0,0,0,0,...,0]
  • ['cinematic','shot']

Here is the prompt with BREAK on different places of the prompt:

Here is the generator used in the images above.

Here is another prompt example from this reddit thread about the BREAK keyword:

The first row starts with the portrait prompt, and the second row, starts with the background prompt. Both are BREAK after each prompt. It also showcases the , before or after or none on the BREAKs.

Here is the generator used in the images above.

AI Model Trigger Words

Based on the dev's comment about the models, the AI model used will be changed based on the keywords/prompt.

Here are some 'normal' keywords that trigger a different model, which will output quite a different result due to the model being used.

• 'Furry' Model
  • Presumably 'Pony Diffusion' Model.
  • my little pony, sonic, chimera, faun, goblin, dryad, mer(maid|man), humanoid, imp, android, cyborg, anthropomorphic, anthro, + any gender, slime girl, aron, twili, centaur, fursuit, paw, avian
  • Any pokemon name(?): abra
  • To increase quality, as per Allo's recommendation, see this article from CivitAI
• 'Anime' Model
  • anime, slime girl, danbooru, pixiv

Sample Prompts

You can visit the following sites for prompt inspirations:

• Stable Diffusion Web - Prompt Search
• CivitAI
• Lexica.art
• SeaArt.ai
• Stable Diffusion Subreddit Wiki

Here is a generator that showcases artists and their effect on the Text to Image Prompt:

• Artist Style Effect Tester - thanks to Allo (see their text-to-image generator here) for the artist list and Ashenthorn for the Artist Plugin and compiling the artists.

Less is More usually, but you can finetune with hundreds of tags/negative prompts.

t2i-framework Plugin

You can use the prompt techniques here on the textarea of the t2i-framework generators like so:

Getting the Input Settings

You can get the input settings upon generation finish like so:

promptSetting
  prompt = [prompt]
  negativePrompt = [negPrompt]
  guidanceScale = 7
  resolution = 512x512
  seed = -1
  size = 400
  onFinish(data) =>
    console.log(data.inputs)
    // You can access the inputted settings 
    // with `data.inputs.prompt`, `data.inputs.seed`, etc.
  

Gallery

This guide will not talk about how to use the gallery, please see the Plugin Page for the settings and how to use the gallery.