A custom extension for AUTOMATIC1111/stable-diffusion-webui that implements an expressive template language for random or combinatorial prompt generation along with features to support deep wildcard directory structures.
Using this script, the prompt:
A {house|apartment|lodge|cottage} in {summer|winter|autumn|spring} by {2$$artist1|artist2|artist3}
Will produce any of the following prompts:
- A house in summer by artist1, artist2
- A lodge in autumn by artist3, artist1
- A cottage in winter by artist2, artist3
- ...
This is especially useful if you are searching for interesting combinations of artists and styles.
You can also pick a random string from a file. Assuming you have the file seasons.txt in WILDCARD_DIR (see below), then:
__seasons__ is coming
Might generate the following:
- Winter is coming
- Spring is coming
- ...
You can also use the same wildcard twice
I love __seasons__ better than __seasons__
- I love Winter better than Summer
- I love Spring better than Spring
More complete documentation can be found here.
Prefer a tutorial? Click here
Need a wildcard library? We've got you covered.
The extension can be installed directly from within the Extensions tab within the Webui
You can also install it manually by running the following command from within the webui directory:
git clone https://github.com/adieyal/sd-dynamic-prompting/ extensions/dynamic-prompts
In the LastBen colab (and perhaps others), you might see an error similar to this: ModuleNotFoundError: No module named 'dynamicprompts'. There seems to be an issue causing extensions that rely on install.py from being installed properly. A workaround is to run this command before you start the webui:
!python -m pip install dynamicprompts
Dynamic Prompts works particularly well with the following scripts:
- randomize if you want to randomise non-prompt parameters like steps, width/height, sample, CFG scale, etc. Both scripts can work concurrently.
- X/Y Plot - setting Dynamic Prompts to combinatorial mode while using X/Y Plot, lets you exhaustively test prompt and paramter variations simultaneously.
Documentation can be found here
In addition to standard wildcard tokens such as __times__ -> times.txt, you can also use globbing to match against multiple files at once.
For instance:
__colors*__ will match any of the following:
- WILDCARD_DIR/colors.txt
- WILDCARD_DIR/colors1.txt
- WILDCARD_DIR/nested/folder/colors1.txt
__light/**/*__ will match:
- WILDCARD_DIR/nested/folder/light/a.txt
- WILDCARD_DIR/nested/folder/light/b.txt
but won't match
- WILDCARD_DIR/nested/folder/dark/a.txt
- WILDCARD_DIR/a.txt
You can also used character ranges [0-9] and [a-z] and single wildcard characters ?. For more examples see this article.
Instead of generating random prompts from a template, combinatorial generation produced every possible prompt from the given string. For example:
I {love|hate} {New York|Chicago} in {June|July|August}
will produce:
- I love New York in June
- I love New York in July
- I love New York in August
- I love Chicago in June
- I love Chicago in July
- I love Chicago in August
- I hate New York in June
- I hate New York in July
- I hate New York in August
- I hate Chicago in June
- I hate Chicago in July
- I hate Chicago in August
If a __wildcard__ is provided, then a new prompt will be produced for every value in the wildcard file. For example:
My favourite season is __seasons__
will produce:
- My favourite season is Summer
- My favourite season is August
- My favourite season is Winter
- My favourite season is Sprint
You also arbitrarily nest combinations inside wildcards and wildcards in combinations.
Combinatorial generation can be useful if you want to create an image for every artist in a file. It can be enabled by checking the Combinatorial generation checkbox in the ui. Note, num batches changes meaning. With random generation, exactly num_batches * batch_size images are created. With combinatorial generation, at most num_batches * batch_size images are created. This upper limit ensures that you don't accidentially create a template that unexpectedly produces thousands of images.
Combinations are not yet supported, i.e. {2$$a|b|c} will treat 2$$a as one of the options instead of selecting two of a, b and c.
The combinatorial batches slider lets you repeat the same set of prompts a number of times with different seeds. The default number of batches is 1.
By default, the Batch count silder of automatic1111 has a maximum value of 100. This can limit the maximum number of generations when using combinatorial generation. You can change the maximum value of this slider by editing ui-config.json and change:
txt2img/Batch count/maximum": 100
to something larger like:
txt2img/Batch count/maximum": 1000
Select this if you want to use the same seed for every generated image. If there are no wildcards then all the images will be identical. It is useful if you want to test the effect of a particular modifier. For example:
A beautiful day at the beach __medium/photography/filmtypes__
That way you can isolate the effect of each film type on a particular scene. Here are some of the results:
Using Gustavosta's MagicPrompt model, automatically generate new prompts from the input. Trained on 80,000 prompts from Lexica.art, it can help give you interesting new prompts on a given subject. Here are some automatically generated variations for "dogs playing football":
dogs playing football, in the streets of a japanese town at night, with people watching in wonder, in the style of studio ghibli and makoto shinkai, highly detailed digital art, trending on artstation
dogs playing football, in the background is a nuclear explosion. photorealism. hq. hyper. realistic. 4 k. award winning.
dogs playing football, in the background is a nuclear explosion. photorealistic. realism. 4 k wideshot. cinematic. unreal engine. artgerm. marc simonetti. jc leyendecker
This is compatible with the wildcard syntax described above.
The first time you use it, the model is downloaded. It is approximately 500mb and so will take some time depending on how fast your connection is. It will also take a few seconds on first activation as the model is loaded into memory. Note, if you're low in VRAM, you might get a Cuda error. My GPU uses less than 8GB by YMMV.
Use the lexica.art API to create random prompts. Useful if you're looking for inspiration, or are simply too lazy to think of your own prompts. When this option is selected, the prompt in the main prompt box is used as a search string. For example, prompt "Mech warrior" might return:
- A large robot stone statue in the middle of a forest by Greg Rutkowski, Sung Choi, Mitchell Mohrhauser, Maciej Kuciara, Johnson Ting, Maxim Verehin, Peter Konig, final fantasy , 8k photorealistic, cinematic lighting, HD, high details, atmospheric,
- a beautiful portrait painting of a ( ( ( cyberpunk ) ) ) armor by simon stalenhag and pascal blanche and alphonse mucha and nekro. in style of digital art. colorful comic, film noirs, symmetry, brush stroke, vibrating colors, hyper detailed. octane render. trending on artstation
- symmetry!! portrait of a robot astronaut, floral! horizon zero dawn machine, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, 8 k
Leaving the prompt box blank returns a list of completely randomly chosen prompts.
This option randomly selects a keyword in your prompt and adds a random amount of emphasis. Below is an example of how this affects the prompt:
a portrait an anthropomorphic panda mage casting a spell, wearing mage robes, landscape in background, cute, dnd character art portrait, by jason felix and peter mohrbacher, cinematic lighting
Tick the Fixed seed checkbox under Advanced options to see how emphasis changes your image without changing seed.
Check the write prompts to file checkbox in order to create a file with all generated prompts. The generated file is a slugified version of the prompt and can be found in the same directory as the generated images, e.g. outputs/txt2img-images
Jinja2 templates is an experimental feature that enables you to define prompts imperatively. This is an advanced feature and is only recommended for users who are comfortable writing scripts.
To enable, open the advanced accordion and select Enable Jinja2 templates.
You can read about them in more detail here
The extension looks for wildcard files in WILDCARD_DIR. The default location is /path/to/stable-diffusion-webui/extensions/sd-dynamic-prompts/wildcards. It can also be manually defined in the main webui config.json under wildcard_dir. When in doubt, the help text for the extension in the webui lists the full path to WILDCARD_DIR
The collections directory contains modifier libraries that you can use as is or to bootstrap your own. To get started, either use the Wildcard Manager tab to copy a one or more collections to your wildcards folder, or you can manually copy the files across. There are currently three collections:
Jumbo is a very large collection of wildcards across many categories including aesthetics, appearance, artists, medium, style, and time. It is a work in progress, but aims to provide good coverage of various modifier categories.
Parrotzone is a far smaller and more manageable collection sourced from https://proximacentaurib.notion.site/e28a4f8d97724f14a784a538b8589e7d?v=42948fd8f45c4d47a0edfc4b78937474.
Devilkkw focuses more on character building, clothes, gestures, food, etc
If you're using a Unix/Linux O/S, you can easily create a symlink to the relevant collection rather than copying it across if you don't plan to alter it. E.g.
ln -sr collections/parrotzone wildcards/
If you're interested in contributing to the development of this extension, here are some features I would like to implement:
- Saved templates - publicprompts.art produces great prompt templates. e.g.
Funky pop Yoda figurine, made of plastic, product studio shot, on a white background, diffused lighting, centered
You can swap out Yoda for anything to create a great character. If I create/discover a great prompt that can be turned into a template, it would be great to have a place to save it. The extension could provide some initial templates. Users could then create and save their own.
-
Tag-based modifers. Creating a strict taxonomy of concepts for modifiers is always going to be flawed. Some modifiers fit in multiple categories. It would be better to introduce tags to make it easy to place modifiers in multiple places at once.
-
Improved modifier management. The modifier library is a little clunky at the moment. You can find a nested hierarchy of all wildcard files, but it isn't possible to see what's in them or even edit them from within the interface. It might be best to implement this in a separate tab.
-
Improve the modifier library. I'm currently creating the library from multiple sources but building taxonomies is hard, especially if you're not a domain expert. The photography section is a good example, there are wildcard files for lighting, filetypes, camera models, perspective, photo websites, etc. Someone with a better understanding of photography might have a better way to divide this. The artists section in particular needs some TLC.
-
Option tweaking. There have been a few requests for option tweaking like XY Plot. I'm not sure that that is in scope for this extension, but it would be nice to see if we could find a solution that leverage's XY Plot.
-
Improved UI. The current UI works, but a few javascript tweaks could improve the user experience tremendously.
Let me know if there is anything you would be interested in looking into.