AI Can Now Help You Write Better Prompts...for AI

You can ask text-wrtiting AI to suggest prompts for your art-generating AI. What's even left for us humans to do anymore?

This is it.

We did it, folks!

We’ve come full circle and are living in a world where we have an AI tool designed to help us create input for another AI tool.

AInception?

Which, according to AI, looks like this.

Remember how I suggested using Lexica.art to get inspiration for your AI text prompts?

Turns out, you no longer even have to do that.

If you just have a vague idea of an image you’d like to create, you can ask an AI tool to suggest a bunch of text prompts automatically.

“MagicPrompt” step by step

Here's the process:

1. You go a Hugging Face tool called Stable Diffusion Prompt Generator (sexy!)

2. You type in your initial text. This could e.g. be your main subject:

3. Then you click “Submit” and let the algorithm do its thing:

4. Boom! You have a number of prompts related to your subject matter that you can test drive in your AI art generator of choice.

How it works

How does it do that?

Well, the MagicPrompt model has itself been trained on the library of images and text prompts from Lexica.art. So behind the scenes, it tries to match your initial text with word combinations used in similar prompts.

You can read a bit more about the model if you’re interested.

Images generated from MagicPrompt’s text prompts

Now let’s see what types of images our four AI-suggested text prompts will create when we pop them into a Stable Diffusion image generator.

I’ll use my go-to dreamlike.art site for this.

Prompt #1

Bicycle on a street corner, Makoto Shinkai, by Artgerm, by beeple, by Greg Rutkowski, volumetric lighting, octane render, 4K resolution, trending on artstation, masterpiece

I actually really like the atmopshere in this image. There’s a story lurking behind it, and all from a simple subject-based starting point.

Prompt #2

Bicycle on a street, very realistic, very realistic, hyperrealistic, highly detailed, very detailed, extremely detailed, detailed, digital art, oil painting, trending on artstation, headshot and bodyshot, detailed face, very detailed face, extremely detailed face, HD Quality, 8k resolution

The prompt here was a bit of a hit-and-miss, with lots of repeated and redundant descriptors as well as odd choices like focusing on a “detailed face” where a bicycle is the main subject.

But the final outcome isn’t too shabby and certainly has a very different vibe compared to our first prompt.

Prompt #3

Bicycle on a street, no wheels, epic wide shot, panorama, 35mm, artstation, artgerm, artstation 3d, artstation render, artstation 3d render, octane render, unreal engine, magical realism, ultrarealistic, chiaroscuro, octane render, 8k, High contrast

There were again some questionable choices in the prompt such as the “no wheels” reference (which Stable Diffusion also chose to ignore in all 8 images it rendered).

But it’s another curious and distinct shot.

Prompt #4

Bicycle on a street, Bizarre, mixed media, digital art, trending on artstation, 8K, epic composition, highly detailed, AAA graphics

Yet another take on the same subject matter.

As you can see, using the prompt generator is a great way to explore different styles while keeping your subject in focus.

Over to you…

How do you come up with your AI text prompts? Do you test them through trial-and-error with free AI tools? Do you browse reference sites like Lexica.art? Do you just describe what you want in natural language and hope for the best?

If you know of a tool or site that makes prompt engineering smarter or more efficient, I’d love to hear about it. Drop me an email or leave a comment below this post.

Thanks for reading!

---