Right now, Stable Diffusion can only take in a short prompt. What if you want to illustrate a full story? Cue Long Stable Diffusion, a pipeline of generative models to do just that with just a bash script!
Start with long-form text that you want accompanying images for, e.g. a story to illustrate.
Ask GPT-3 for several illustration ideas for beginning, middle, end, via the OpenAI API.
“Translate” the ideas from English to “prompt-English”, e.g. add suffixes like trending on art station for better results.
The “prompt-English” prompts are put through Stable Diffusion to generate the images.
All the images and prompts are dumped into a .docx, for easy copy-pasting.
Purpose
I made this to automate my self, ie. prompt AI for illustrations to accompany AI-generated stories, for the Stories by AI podcast. Come check us out! And please suggest ways to improve—comments and pull requests are always welcome 😃
This was also just a weekend hackathon project to reward myself for doing a lot of work the past couple of months, and for feeling guilty about not using my wonderful and beautiful Titan RTXs to their full potential.
Run
This bash script runs what you need. It assumes 2 GPUs with 24GB memory each. See the note above, under Steps, to change this assumption for your compute needs. I had too much fun with multiprocessing and making it faster.
bash run.sh -f three_little_pigs
To run your own text, replace three_little_pigs with the name of your new .txt file, put in the texts/ folder.
bash run.sh -f <name_of_txtfile_in_texts_dir>
What you need before you run it like that
Install the requirements
Make sure you set your OpenAI API key, e.g. in terminal export OPENAI_TOKEN=<your_token>
Make sure you have run ‘huggingface-cli login’ with a valid token
For using extracts method, you need to install nltk and run nltk.download(‘punkt’) in a python shell
Then, put your favorite story or article in a .txt file in the texts/ folder
Method Selection
Currently two methods for generating the image prompts from text are supported.
“sections”: Inputs the entire text to GPT-3 and tells it to generate images for the start, middle, or end of the text.
“extracts”: Splits the text from the .txt file into smaller chronological bits of text, and then generates an image prompt for each bit of text.
Additional methods yet to be implemented are following:
“summary”: Generates a summary from the .txt. file, then prompts GPT-3 to generate image prompts from the summary.
“summary+extracts”: A combation of “summary” and “extracts” method, where both summary and the extract are fed into GPT-3 to generate image prompts.
Output Selection
Currently one type of output is supported
“docx”: A word file with the images and prompts.
Additional output formats yet to be implemented are:
“txt”: Just a text file with the image prompts (does not run stable diffusion).
“images”: Just image PNG files with their title being the prompt.
“html”: A self contained HTML page with the original text and suggested images
“markdown”: A markdown file with the original text and image embeds
“latex”: A latex file with the original text and ‘’ components for the images
“pdf”: A self contained PDF documents with the original text and images, compiled from latex
Files and folders
run_two_gpus.sh: This is the main entry script into the program to parallelize across GPUs easily.
run.py: Where most of the magic happens: getting image prompts from GPT-3, making images from those prompts (using stable diffusion, multithreading), saving all those and also dumping those images and prompts to a docx file. This is what run_two_gpus.sh calls.
stable_diffusion.py: Just runs stable diffusion if you want to use it by itself (I do). run.py calls it.
dump_docx.py: Just dumps image prompts and images into a single docx for a particular text. Again, it’s useful if you want to use it by itself on the saved images and prompts. I do, because I’m actually overwriting the file when multiprocessing and sometimes will just use this as a postprocessing step. Yes, you can join those and change that but I don’t really care, since sometimes my GPUs misbehave and I’ll need to rerun it anyways.
texts/: Folder to put your texts in, as a .txt file.
image_prompts/: Generated image prompts by GPT-3 based on your text.
images: Generated images by Stable Diffusion based on GPT-3’s image prompts.
docx/: Microsoft Word document for a text with images and their prompts all in one.
clean_lexica.py: Preprocessing step for Stable Diffusion prompts from Lexica - clean up the prompts and put them into a single file.
effective_prompts_fs.txt: Effective “prompt-English” to use for few-shot translation from English GPT-3 prompts to prompt-English (1884 tokens).
Multi-processing Multi-GPU Note
Multi-processing is optimized for 2 Titan RTXs, with 24GB RAM each. Changing the number of GPUs to parallelize on is a simple edit in run_two_gpus.sh: just copy the first line and change CUDA_VISIBLE_DEVICES to the appropriate GPU id.
Changing the number of processes for each GPU is an argument that can be passed in through run_two_gpus.sh as -n <num_processes_per_gpu> for each run. This is an int used in run.py. I’ve found that my GPUs can handle 3, but are happier with 2.
Complete
[x] Pipeline of asking GPT3 for image prompts
[x] Image prompts to stable diffusion
[x] Multiprocessing to max out a single GPU
[x] GPU multiprocessing stable diffusion
[x] Docx dump of images and image prompts
[x] Translation layer between English prompt and “prompt English” (lexica)
[x] Flesh out readme
[x] Open source
Todo
[ ] Walkthrough video of code
[ ] Replace stable_diffusion.py with txt2img.py from CompViz stable-diffusion repo
[ ] Support for configuring image generation (based on txt2img.py)
[ ] Support for different content types (fiction/blog post/essay/news article)
[ ] ‘summary+extract’ method
[ ] output to txt
[ ] output to markdown
[ ] output to markdown
[ ] output to html
[ ] output to latex
[ ] output to pdf
[ ] refactor from a sequence of script to a python library
Future
[ ] Translation from English to ‘prompt English’ can be improved with: finetuned model with several million data samples (instead of 36)
