Opensource prompts for arible.co, precursor to the arible format
Simdi:
Hey, have you seen this idea about a Universal Prompt Format for AI-generated images? It aims to allow reproducing the same image exactly by including all the necessary information in a single file.
P1:
Oh, that sounds interesting! How would it work?
Simdi:
The idea is to include the model details, training parameters, control-net based parameters, depth and pose information, and in-painting masks in a single file. So, instead of asking the OP about the steps they took, you'd just ask for the file and import it into your web UI to get the same image.
P1:
That's a neat concept. It could be helpful for professionals or even for teaching and learning purposes. But I have some concerns about including the model and training images, as they might be too big to share in a single file.
Simdi:
You're right, including the entire model and training images might not be feasible. We could just include the model name and hash instead, and expect users to have the model as a dependency.
P1:
How do you envision tracking and incorporating inpainting operations within the Universal Prompt Format, given that they can be run multiple times on a single image with different masks, prompts, and parameters?
Simdi:
It might be difficult but not impossible. We'd need adoption by or cooperation with various UIs to keep a history of inpainting operations, adding to the file's history every time there's img2img or inpainting.
P1:
How could the Universal Prompt Format incorporate version control to track and manage changes made to the AI-generated images, especially when working with different models and user inputs?
Simdi:
The Universal Prompt Format could incorporate version control by using a system similar to Git. Each change made to the AI-generated image would create a new version, tracking changes and allowing users to revert to previous versions if needed. The system would store metadata about the changes, supporting branching and merging for collaboration. This approach would simplify managing and reviewing changes when working with different models and user inputs.
Follow me on twitter: @nintharc for updates