README.md

metadata

license: mit

RandomPrompt-v1

A fine tuned GPT-neo 125M

The purpose of this model is to autocomplete or generate danbooru-like prompts for generating images in Stable Diffusion derivatives that use danbooru tags for text conditioning.

Usage

THE HOSTED INTERFACE DOES NOT WORK, USE THE HUGGINGFACE SPACE

Autocompletion

Type in a few tags, and it will generate a completion of the prompt

Generation

Type in nothing, and it will generate a prompt

Training

Trained on 400k tags from danbooru posts for 600k steps, or around 0.25 epochs

https://wandb.ai/saltacc/RandomPrompt/runs/2v2arf0u?workspace=user-saltacc

I plan on doing further runs on better hardware to try to get more accurate prompt completion