LLM-Based Prompt Brain Storming

Link to interactive prompt tree here.

Warning: This link includes NSFW images.

I wanted to run a little experiment to add more more variance to the process of doing human evaluations of models. Running a model on the same set of prompts is good for consistency and direct comparison - but sometimes it's refreshing to ... just explore, ya know?

So the idea is to essentially do prompt brain storming and visualize the results in some interactive way.

We'll use a simple prompt:

PROMPT_TEMPLATE = """
Rewrite the provided image prompt.
You should {variation}.

Your prompt should favor using danbooru tags in the prompt, but include natural language when necessary.
Keep the prompt short and concise, less is more. Less than 20 words.

Answer only with the rewritten prompt - no preamble or postamble.
Include no additional comments or explanations.
Your reply should be one single prompt, on one line. No newlines.
Don't use parentheses or brackets in your prompt.
Don't put the prompt in quotes.

PROMPT: "{prompt}"
"""

We then introduce two variations:

    for variation in [
        "extend the idea, add something, alter the style or otherwise evolve the prompt",
        "take a creative leap and reframe the entire prompt in a new and interesting way which goes in a new and interesting direction"
    ]:

We could introduce more variations, but we have to trade depth for breadth. Since what we will essentially do is start at a root prompt and recurse this process for all varied prompts.

This creates a growing tree

At each prompt node we get two variations

You can explore the interactive tree here. Warning: This link includes some NSFW images.

The same tree can be run with multiple models to compare results, like one would expect, but it's also quite fun to generate new trees!

The LLM I used is Sao10K/L3-8B-Stheno-v3.2, which I find handles this task quite well. But it's a bit spicy and uncensored - so be aware.

Something to play around with is to add randomness to the "forking" at the hands of variations, if we for example only do two variations 10% of the time, we can afford to go deeper.

LLM-Based Prompt Brain Storming

Comments