Full Description

Kidechisms! is an AI-powered spiritual formation tool for kids. It takes the 196 questions and answers from the Westminster Larger Catechism and autogenerates: 1) an entertaining and contextualized story with illustrations 2) graphic stickers for child interaction 3) in the future, it will generate a catechism-based prayer that the parent can say with the child (we didn’t have time to build this feature). This product makes it easy for parents to engage with their children in spiritual formation and discipleship since it also makes it fun for the kids. Consider using it for bedtime stories! Underneath the hood, we are addressing the difficult task of applying text-to-image AI to create consistent characters across scenes for our story-telling application. You cannot do this by mere prompt engineering. To solve this, we used a LLM to parse scene descriptions for creating backgrounds and a two-stage inference procedure with latent diffusion models to create characters that can be placed in the scene. The two-stage procedure uses a fine-tuned version (using LORA) of stable diffusion XL to generate base sticker characters and a text+image to image generator with Euler Ancestral Sampling to steer the generation of the same character with different expressions. We implemented this using the Hugging Face Diffusers library. Our image generation can run on consumer GPUs that have 24GB of ram. The impact of solving text to image is enormous. It allows for all sorts of innovations in multimodal content generation for a wide range of useful applications.