?? Text to Image Plugin ??

This plugin allows you input some text and get an image out. It uses Stable Diffusion, an AI image-generation model. It doesn't run on your actual device like other Perchance plugins because it requires too much computational power (and would require a 3GB download), so it runs on server GPUs, which means it costs me money to run. For this reason, this plugin is funded with ads, so an ad will appear on your generator for non-logged-in users if you import this plugin. The ad will appear at the bottom of the screen like this. The ad will go away if you remove the plugin, of course. Please see the notes at the end of this page for more info.

To use this plugin, you'll first need to import it by adding this line to your lists editor:

image = {import:text-to-image-plugin}

And now try putting this in your lists editor:

character
  a {mech|demon|cyberpunk} {warrior|minion|samurai}

place
  soviet russia
  a small village
  a mountainous region
  an underwater cavern

season
  winter
  summer
  
prompt
  detailed painting of [character] in [place], [season]
  
output
  [image(prompt)]

Now just write [output] in the HTML wherever you want an image to appear. Here's a live, working example of what that outputs:

randomize

You can hover your mouse over the image (or long-press on mobile) to see the prompt that was used, or click the info icon in the corner of the image. You can also manually display it below the image by using the special lastTextToImagePrompt variable:

output
  [image(prompt)] <br> [lastTextToImagePrompt]

Here's an example generator that uses the above code. Try playing around with the lists and saving your own copy.

As the name suggests, [lastTextToImagePrompt] will always contain the most recently used prompt. If you instead wrote [image(prompt)] � [prompt] then the prompt used to generate the image (seen on hover) and the prompt output under the image would be different, because each time prompt is evaluated, it is randomized (it's just a normal Perchance list, after all).

If you want the prompt text to be above/before the image, you can do that like this:

output
  [p = prompt.evaluateItem] <br> [image(p)]

Here's an example generator that uses the above code. And this example shows how to add multiple images to your generator.

Here's an example generator that has multiple images and also allows the user to input a text prompt.

There are some options/settings that you can set two different ways - the first is by putting them in a promptData list like this:

promptData
  prompt = painting of [character] in [place], [season]
  seed = 123
  size = 400  // size is only a valid property for square resolutions
  style = border:4px solid blue; margin-top:20px; // CSS styles

You'd then write [image(promptData)] to generate an image using those settings (and in this case you can use [promptData.lastUsedPrompt] instead of [lastTextToImagePrompt] to get the prompt that was used if you want).

The second way is to put the options directly in your prompt text like this:

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

Here's an example generator that has the options/settings within the prompt text itself. The options should always be at the end of the prompt, and should follow the (name:::value) format.

You can of course omit settings that you don't want to customize.

You can choose between 3 different resolutions using the resolution. The valid resolution values are 512x512, 512x768 and 768x512:

promptData
  prompt = fantasy {forest|city|village|cafe|cavern|island|plains|castle|canyon|supercity|megalopolis}, extremely detailed oil painting, unreal 5 render, rhads, bruce pennington, studio ghibli, tim hildebrandt, digital art, octane render, beautiful composition, trending on artstation, award-winning photograph, masterpiece
  resolution = 512x768
  width = 400  // height will be auto-chosen based on aspect ratio if omitted, and vice versa for width

Here's an example generator that uses the above code, and here's a live demo of that:

randomize

There are a couple of other parameters to play with:

• negativePrompt: Tell the AI what you don't want in the image. E.g. if you don't want any blurriness in the output image, you'd write something like negativePrompt = blur, blurry image, motion blur. Here's an example generator showing this feature.
• guidanceScale: Roughly speaking, this controls how much the output image "matches" the prompt. You can make the value higher to make the output "match" the prompt more, at the expense of realism. The default value is 7, the minimum is 1, and the maximum is 30.

You'll notice that when you hover your mouse over the image there's a button which opens a menu that allows you to save images to a public gallery (for your generator), and to display said gallery. You can set the title and description that a gallery image will be saved with like this:

promptData
  prompt = ...
  saveTitle = ...
  saveDescription = ...

If you don't set a saveTitle and saveDescription, then by default the title will be the part of the prompt that comes before the first full-stop/comma/question-mark/exclamation-mark, and the description will be the whole prompt.

If you'd like to display the gallery on your page, rather than users having to click the button to open it, you can use "special" options list with the gallery property like this:

galleryOptions
  gallery = true
  sort = top // or 'recent' or 'trending'
  timeRange = 1-week
  hideIfScoreIsBelow = -2 // images will be removed if they get down-voted to a score below -2
  adaptiveHeight = true // expand height to fit all images (so there's no scrollbar on the gallery)
  contentFilter = pg13 // or 'none' - EDIT: This is now always set to pg13 by default. The user must choose to switch to filter-less mode themselves.
  style = ... // optional CSS styles (you can delete this line)

And then just put this in your HTML editor (bottom-right editor):

[image(galleryOptions)]

The valid values for timeRange are: 1-day, 3-day, 1-week, 1-month, 1-year, all-time. Here's an example generator that displays the gallery.

You can ban users and prompt phrases from the gallery using the bannedUsers, bannedPromptPhrases, and bannedNegativePromptPhrases options. Have a look at this example to see these features in action.

galleryOptions
  gallery = true
  // ...
  bannedUsers // click the settings button at the top of the gallery and type "admin" to toggle admin mode on, then double-click on an image to get the user ID of the creator.
    263efb15c47c2d2f398e91bf169f50d4a0ca69251638c9d0eb5823c0e4fba538
    f50d4a0ca69251638c9d0eb5823c0e4fba538263efb15c47c2d2f398e91bf169
  bannedPromptPhrases
    pg13:blood // ban the word 'blood' in pg13 mode
    /twin.?towers?/ // example of 'regex'-based pattern matching to ban 'twin towers' or 'twin-tower' or 'twin_towers', and so on
    pg13:/\b(gore|blood)\b/i // another example of 'regex'-based pattern matching - uses word boundaries and case-insensitive matching
  bannedNegativePromptPhrases
    pg13:wearing clothes // ban the word 'wearing clothes' in the *negative* prompt when in pg13 mode

You can click the settings button at the top of the gallery and type "admin" to toggle on "admin mode". This will show images that contain banned phrases with a red border instead of hiding them (useful for debugging regexes and ensuring that your ban lists aren't banning harmless prompts), and you can double-click on any image to get the user ID of the creator. Again, look at this example, for an example of these moderation features.

Notes:

• You can use this example to get started. And here's another that hides the irrelevant parts of the prompt from the user.
• Images are not stored on the server unless the user explicitely saves them to the gallery - see this post for more info.
• If you want to programmatically get the actual image data that is generated - so e.g. you can draw some text on it, or make it greyscale, or collage multiple images together, or whatever, check out this example.
• The quality of the output image can change dramatically depending on the wording in your prompt. You can use a generator like this to play around with your prompt design (click the info icon on the output images to see the full prompt used).
• You can call the promptData list whatever you want. If your settings list was called promptSettings then you'd write [image(promptSettings)] to generate the output image. You can have many prompt-settings lists in one generator.
• The seed parameter should be any number like 3834329 or 9278236492. A seed of -1 is default and means "choose a random seed for me". If you provide the same seed with the same prompt, it'll generate exactly the same picture. But note: I'll be upgrading the machine learning models that power this as new ones are released, and during the upgrades, the image that a seed+prompt combination "refers to" will change.
• If a seed of -1 is used (which again, is default), then an icon will appear (when you hover over the image) to allow you to try generating it again to get a different result. If you want to add your own "try again" button that just regenerates the image and nothing else, then add id=yourImageId to your promptData list and then use this code to create your "try again" button: <button onclick="yourImageId.reload()">try again</button>. Here's an example generator that does that.
• The model can return NSFW/adult-themed results if prompted with NSFW/adult-themed terms. Treat this like a Google image search, and prompt responsibly. You can add terms like "NSFW" and "nudity" to the negativePrompt option as a way to reduce the probability that you'll get accidental NSFW results. May also want to add "fully clothed" to the prompt in some cases.
• Stable Diffusion was trained on captioned internet images, so you should try adding words to your text prompt that would appear in the captions of high-quality images online like "trending on artstation" or "highly-detailed" or "by Caspar David Friedrich" (a famous artist). You can look up more prompting tips for Stable Diffusion online. This example uses a few common modifiers. An important thing to note is that the machine learning model that currently powers this plugin (that is, Stable Diffusion) will likely be upgraded to a better version in the future, so you should try not to use prompting techniques that are very specific to Stable Diffusion.
• Each user can only have a few concurrent server requests, so if you have lots of images on one page, they'll queue up.
• By using this plugin, you agree to the Stable Diffusion Terms of Use.
• The 19th day of every month is observed as 'Ad-viewer Appreciation Day' in the Perchance community. On this day we pay our respects to the non-logged-in users who fund the GPU servers by viewing ads on generators that import AI-based plugins. Logged-in users are encouraged to spare a moment for these anonymous benefactors, wishing for them a month of relevant and interesting ads, and thanking them for their tolerance of increased browser tab memory usage, and their indirect but valuable contribution to the Perchance community via the digital attention economy. May their mobile game ads not be too sus, and may the gameplay reflect the real gameplay even if only abstractly ???
• Check out more plugins at perchance.org/plugins

As some inspiration, here are some images produced using the prompt "fantasy [thing], extremely detailed oil painting, unreal 5 render, rhads, bruce pennington, studio ghibli, tim hildebrandt, digital art, octane render, beautiful composition, trending on artstation, award-winning photograph, masterpiece":