Implementing and testing a “poor man’s prompt expansion” model for Stable Diffusion

Various Stable Diffusion models massively benefit from verbose prompt descriptions that contain a variety of additional descriptors. Much recent research has gone into training text generation models for expanding existing Stable Diffusion prompts with relevant and context appropriate descriptors.

Since it isn’t feasible to run LLMs and text generation models inside most users’ web browsers at this time, I present my “Poor Man’s Prompt Expansion Model“. It uses a number of examples I’ve acquired from Fooocus and Hugging Face to generate completely random (and absolutely not context appropriate) prompt expansions.

(For those interested in following along at home, you can checkout the gist for this script on GitHub).

How does it work?

We iterate through a list of an absolute crap ton of prompt descriptors that I’ve sourced from other (smarter) systems that tokenize user prompts and attempt to come up with context appropriate responses. We’re not going to do that, because we’re going to go into full chaos mode:

  1. Iterate through a list of source material and split up everything separated by a comma.
  2. Add the resulting list to a new 1-dimensional array.
  3. Now, build a new descriptive prompt by looping through the list until we get a random string of descriptors that are between 175 and 220 characters long.
  4. Once that’s done, return the result to the user.
  5. Create a new prompt.

For our experiment, we’re going to lock all image generation parameters and seed, so we theoretically get the same image given the exact same parameters.

Ready?

Here is our base prompt and the result:

Happy penguins having a beer

Not bad! Now, let’s go full chaos mode with a new prompt using the above rules and check out the result:

Happy penguins having a beer, silent, 4K UHD image, 8k, professional photography, clouds, gold, dramatic light, cinematic lighting, creative, pretty, artstation, award winning, pure, trending on artstation, airbrush, cgsociety, glowing

That’s fun! (I’m not sure what the “silent” descriptor means, but hey!) Let’s try another:

Happy penguins having a beer, 8k, redshift, illuminated, clear, elegant, creative, black and white, masterpiece, great power, pinterest, photorealistic, award winning, vray, enchanted, complex, excellent composition, beautiful composition

I think we just created an advertisement for a new type of beverage! It nailed the “black and white”, though I’m not sure how that penguin turned into a bottle. What else can we make?

Happy penguins having a beer, volumetric lighting, Digital, intricate, awesome, futuristic, cartoon artstyle, vector, solid, detailed, dramatic light, realistic photograph, wonderful colors, dramatic atmosphere

The dude in the middle is planning on having a good night. Definitely some “wonderful colors”. Not so much realistic photo or vector, but fun! One last try:

Happy penguins having a beer, 35mm, surreal, amazing, Trending on Artstation HQ, matte painting hyperrealistic, full focus, very inspirational, pixta.jp, aesthetic, 8k, black and white, reflected on the matrix studio background, awesome

As you can see, you can get a wide variety of image styles by simply mixing a bunch of descriptive elements to an image prompt.

I’ve wanted to implement a feature like this on ArtBot for a long time. (Essentially, if the user allows it, automatically append these descriptions behind the scenes when an image is requested). Perhaps this will come soon.

Leave a Reply

Your email address will not be published. Required fields are marked *