jlowin / aimages

Generate images with hidden text

AImages

A script for generating images that contain hidden text.

Blog post and motivation: https://www.factsmachine.ai/p/hidden-in-plain-sight

Setup

Modal

This script is designed to run on Modal GPUs. Before running it, please set up your Modal account from your local machine. Please note that Modal is a commercial product

To run the script, clone this repo to your local machine and install the dependencies:

git clone https://github.com/jlowin/aimages
cd aimages
pip install -r requirements.txt

Running

You'll need to create a control image. For best results, use a PNG of white text on a black background, like this one:

Run the script with Modal, replacing the path and prompt appropriately:

modal run sd_controlnet.py \
    --control-image-path path/to/your/control/image \
    --prompt "a prompt to generate the full image" 

This will save an image on your computer.

Note: the first time you run this script, it will download all the model weights and create a Docker image for you. This can take a few minutes, but only needs to be done once. Subsequent runs should start in less than a minute and take only a few seconds to generate an image.

Options

In addition to the control image and prompt, you can provide many options to control image generation:

Setting	Default	Notes
--seed		Provide an integer seed value to "fix" the model rather than using a random starting point. The seed for every model is always printed to the logs for reference.
--steps	30	The number of model iterations to run.
--n	1	The number of images to generate.
--negative-prompt		Generated images will try to "avoid" this description
--guidance-scale	7.5	The strength of the prompt for the Stable Diffusion model. Values between 7-9 are generally recommended; higher values will attempt to represent prompts more literally.
--controlnet-conditioning-scale	1.3	The strength of the ControlNet. Values between 1-2 are effective; higher values make text more visible.
--sd-model	Lykon/AbsoluteReality	This script automatically downloads a few popular Stable Diffusion weights: runwayml/stable-diffusion-v1-5 (the base SD model), as well as SG161222/Realistic_Vision_V5.1_noVAE, Lykon/AbsoluteReality, and Lykon/DreamShaper. You can choose any of them to adjust the style of your images.