Making Mazes with AI - One Year Later

A year ago I did a long series of posts testing out text to image AI and seeing how they did generating mazes based on 10 prompts. Today I wanted to come back a year later and see how the AI image generation has changed. Rather than re-test all 12 websites, I decided to look at the top 4 finishers from last time, and retest them on 3 of the same prompts (I had used 10 prompts last year). Before we get into the comparisons here is a summary of my conclusions from last year:

COMPARISON OF 12 AI GENERATING WEBSITES - WHO DID MAZES THE BEST ?

WHAT I LEARNED USING AI TO MAKE MAZE ART

Here are the 4 top finishers last year:

starryai logo
Bing Image creator logo
Stable Diffusion logo
Nightcafe logo

Prompt 1 - Make a medium difficulty maze of the Eiffel Tower in black and white with arrows at the start and finish.

In 2023 this was the best image from starryai:

Starryai Eifel Tower 2023

And here are the 4 options this year (Bing allows you to download all 4 options in one photo):

Result analysis: We didn’t get any real mazes again but all 4 of Bing’s options are solid and closest. They all look like the Eiffel Tower (same as last year) and I like that the drone level images include appropriate buildings for Paris.

Prompt 2 (#4 LY) - Draw a difficult maze that looks like a drawing of a famous building in sketch style

In 2023 this was the best image from Night Cafe:

nightcafe famous building

And here are the 4 options this year:

Result analysis: Bing and Nightcafe both allow 4 images at once - and that’s great in this instance because those generations are strong and similar to each other. I like them a lot. Stable Diffusion is clean and starryai is interesting. Overall these are really good even if the mazes aren’t solvable

Prompt 3 (#9 LY) - Make a solvable maze that is very large and very difficult to solve because it is so complex

starryai also won this in 2023:

And here are the 4 options this year:

Result analysis: I like what Bing did even if got a bit weird, Night Cafe has nice concepts and Stable Diffusion looks cool. Every starryai option looked 8 bit Nintendo game-ish.

Overall thoughts:

Did AI image generation improve in a year ? Yes, it did. I have been playing around more with them as a result of this blog post and I believe they have, but not evenly. Starryai doesn’t appear to have improved much to me. Bing and Night Cafe made bigger leaps and I especially like Bing for its consistent quality.

Can it make mazes ? Still no. I’m sure it will in the future since there are plenty of websites out there that can do it already. Eventually AI will read these sites and be able to maze makes: A Comparison of the Top 12 Online Maze Generation Websites. Then the new question will be can it make maze art ? and not just a simple maze !

Making mazes with AI: Google Gemini

Making Maze Art with google gemini

You can access the website here. You must sign into your google account to use Gemini. There is not limit to the number of images you can generate. Each prompt returns 2 images with the option to generate more. For this exercise I chose the one closest to what I had asked for, or the most interesting.

My original testing took place in March 2023. Gemini (formerly Bard) just started this feature in Feb 2024 so it late to the party. Next week I will be going back and seeing how AI image generation has improved in a year:

Making Mazes with AI - one year later

But, next week is next week, let’s get to how Google performs:

Prompt 1 - Make a medium difficulty maze of the Eiffel Tower in black and white with arrows at the start and finish

That is the Eiffel Tower. There is a maze looking section and an arrow. Typically I do not get arrows, so this is an improvement.


Prompt 2 - Draw a medium difficulty large maze of the Empire State Building with the start and goal embedded in the structure

Similar to the 1st maze but in a new style. Got the famous building correct. Looks like a maze but is not a maze.

Prompt 3 - Draw a difficult maze of the White House pixel art style

Again, we have the right building. Wrong style. No Maze. But good job if I had written a different prompt.

Prompt 4 - Draw a difficult maze that looks like a drawing of a famous building in sketch style

The top is a sketch of the very famous ______ building. The bottom is a hand drawn maze that actually is a maze, i mean if it had a start and goal !

Prompt 5 - Draw a maze in the style of doyoumaze.com of a skyscraper in NYC

Cool NYC image. I think I like the marble sky. No Maze, not my style, but at least it is cool !

Prompt 6 - Draw a maze in the style of Sean C Jackson of a scene from a large outdoor market

This has potential. Change the perspective and switch from watercolor to more comic and you’ve got it.

Prompt 7 - Make a maze of a slice of an orange in color

Sure.

Prompt 8 - Make a maze integrated on top of a photograph of a king sitting on his throne looking cantankerous beside his beautiful queen

Weird. I didn’t ask for a specific king and queen. Any would do. If I ask for a plumber or bus driver is it blocked ? And, after I asked this…well you all saw the many stories !

Prompt 9 - Make a solvable maze that is very large and very difficult to solve because it is so complex

Looks like a hedge maze from above. Nice style.

Prompt 10 - Make a 3d render of a red and blue glossy cube maze

No.

How did Google Gemini do ? I think it was average. It has less structure than other text to image generators. You can ask for whatever you want (except people I guess). And there are no limits. But it also has no built in styles or adjustment buttons that are helpful with other AI image sites.

So this post comes many months after my initial AI review:

Comparison of 12 AI generating websites - Who did mazes the best ?

Would any of these Google Gemini images make the best of ? I think the first 2 would be considered. The orange slice (not the one you want to win) and maybe prompts 6/9. Still the overall is average. I still prefer my original picks.

What I learned using AI to make maze art

When I started this blog series on AI image generators I said I wanted to answer some questions. After spending a full month using and working with a dozen text to image generators let’s see what I have learned. Here is the original post with the questions, and my added answers:

What do I hope to accomplish ?

  • See what the competition is. How good is the output.

The output was mixed, with some generators producing intriguing abstract patterns while others struggled with basic maze structures. I did not find any generator who could actually make maze art if only because the current text to image generators can’t make a solvable maze. They can make maze and labyrinth inspired images that are wonderful. They just can’t be solved. I think this will be fixed just like “hands” were fixed.

  • Become more familiar with what AI image generators are capable of

It seems the standard is to generate 4 images from every prompt text you input, allowing you to pick the best. Many of them also have pre-selectable styles. Each site has a slightly different style and strengths. I found the sites I like to use and those I don’t.

  • Evaluate and compare the different sites to determine how well they create based on my prompts

As discussed in my previous post there are 4 sites that were better than the rest in my challenge ! Different prompts may mean slightly different results but I imagine the order would not meaningfully change much with new prompts. Some sites just work better than others.

  • Be inspired (hopefully)

Yes. So many interesting takes on my text prompts (many more terrible ones, but..). I think these sites are a good place to look for inspiration.

  • Answer: Can I compete with AI generated maze art ?

Right now, they cannot compete with the current quality of maze artists. The odd thing is that a standard program for computer programmers to write is an algorithm to create a maze. There are dozens of sites, many that I have reviewed before to do this. At some point an AI text to image will learn from these sites and I will need to come back to reevaluate this.

And some additional things I learned:

  • Asking for a famous building means you may not get the full building - usually the top is cut off for some reason. Here is a gallery of the prompt “Make a medium difficulty maze of the Eiffel tower in black and white with arrows at the start and finish” from many different sites. I think none of these actually came out as I expected and none of them were able to execute the prompt.

  • Prompts are often ignored. Expect the AI’s to miss some things. It might be you and a bad prompt…it might be the AI. A few sites allow you to select an image to “start” with as inspiration. So if things aren’t quite what you want, keep manipulating an image until you get it right.

  • Sometimes you write a bad prompt. If you keep getting bad outputs on multiple websites, the problem is your writing. I would give the example I used for #7 “Make a maze of a slice of an orange in color” didn’t work. My bad.

  • It's best to approach AI image generation with an open mind, rather than trying to achieve a specific outcome. Better to try a prompt and see what comes out. I cannot tell you how many times I have prompted “make a maze on the surface of the Death Star” because I knew it would be awesome and I have yet to be impressed. Here is one from StarryAI. It looks cool but is not what I wanted because I was expecting a specific look.

AI - make a maze on the surface of the Death Star

While this is the end of this blog series, I will continue to play around with AI generators and you can expect more content from me on the subject !

Conclusion:

While AI image generators can create visually interesting maze-like patterns, they still fall short of producing truly solvable mazes.

If you missed any of the previous 13 blog posts in the series here is the full series with links:

An exploration into Al Image Maze Generation

Making mazes with AI: DallE2

Making mazes with AI: Stable Diffusion

Making mazes with AI: Craiyon

Making mazes with AI: Dream by Wombo

Making mazes with AI: Nightcafe

Making mazes with AI: StarryAI

Making mazes with AI: DeepAI

Making mazes with AI: Canva

Making mazes with AI: AI Image Generator

Making mazes with AI: ToolBaz

Making mazes with AI: Photosonic

Making mazes with AI: Bing AI Maze Generator

Comparison of 12 AI generating websites - Who did mazes the best ?

See you soon with more maze content ! And don’t forget to check out my maze art !

Can AI Generate Mazes? We Tested 12 AI Art Generators to Find Out

Welcome to the 14th in a series of posts where I will test AI image generators and see how they handle making maze art. I have asked the same 10 prompts from 12 different websites. Today we compare the best from each prompt and see who does AI text to maze image the best. If you need to catch-up here is the series so far:

An exploration into Al Image Maze Generation

Making mazes with AI: DallE2

Making mazes with AI: Stable Diffusion

Making mazes with AI: Craiyon

Making mazes with AI: Dream by Wombo

Making mazes with AI: Nightcafe

Making mazes with AI: StarryAI

Making mazes with AI: DeepAI

Making mazes with AI: Canva

Making mazes with AI: AI Image Generator

Making mazes with AI: ToolBaz

Making mazes with AI: Photosonic

Making mazes with AI: Bing AI Maze Generator

Making Maze Art with AI

Welcome to the finals of my maze making AI challenge. For the finals I will pick the 4 best images from the 12 sites. It is subjective, but there are a few that we could all agree do not make the cut. I will pick 4 finalists, then rank them. At the end I will rank all the sites based on how they did. You can do your own ranking if you would like.

Prompt 1 - Make a medium difficulty maze of the Eiffel Tower in black and white with arrows at the start and finish.

starryAI - Eiffel Tower Maze
Photosonic - Eiffel Tower Maze
Bing AI - Eiffel Tower Maze

The finalists are Stable Diffusion; StarryAI; Photosonic, and Bing AI. I should say that all of the generations were good for this prompt. My rank 1 - StarryAI wins for having the needle included and a maze and the cool viewpoint. 2 - Bing AI a strong second with the interesting maze, 3 - Stable Diffusion, 4 - Photosonic.


Prompt 2 - Draw a medium difficulty large maze of the Empire State Building with the start and goal embedded in the structure

There was a real struggle showing the needle of the building in these generations. Also, including a maze seemed to be difficult for the AI. Our finalists are StarryAI, Nightcafe, Stable Diffusion, and BingAI. My order 1 - Stable Diffusion (because it has the needle and the maze), 2 Bing AI - very cool concept ! 3 - StarryAI because at least it looks mazelike in black and white, and 4 - Nightcafe.

Prompt 3 - Draw a difficult maze of the White House pixel art style

Most AI could do pixel style and my finalists did it best. If I had said make a grid instead of a maze these would be much better. My finalists are Photosonic, Stable Diffusion, Dream by Wombo and Canva. I order them 1 - Stable Diffusion, 2 - Photosonic, and 3 - Dream by Wombo who almost got the building right and correctly included the flag…but that yard ! 4 - Canva who did a 3/4 of a great white house but forgot the maze.

Prompt 4 - Draw a difficult maze that looks like a drawing of a famous building in sketch style

4 different looking finalists. Canva, Nightcafe, Stable Diffusion and StarryAI made the grade. I want to give a special shoutout to StarryAI who had 4 different outputs that were all very interesting. As far as the other AI generators they typically made a simple maze, some with doors to signify a building. Not very good. My order is 1 - Nightcafe for the great fantasy looking castle in color, 2 - Stable Diffusion, because it has a maze in front of a column building and 3 - Canva, cool building in black and white that looks kind of famous. 4- StarryAI - love it, but doesn’t look like any famous building I have ever seen.

Prompt 5 - Draw a maze in the style of doyoumaze.com of a skyscraper in NYC

None of these used my style (that is good news for me !). Finalists are Dream by Wombo, StarryAI DeepAI, Stable Diffusion. I think the first 2 are very close…I think I will make the order as it is shown.

Prompt 6 - Draw a maze in the style of Sean C Jackson of a scene from a large outdoor market

I want to say that the output of this prompt was varied and interesting. I went with a more famous maze artist to test if the AI would get it (they generally did not). Finalists are Canva, Dream by Wombo, Nightcafe and Bing AI. Because it was closest to what I asked for 1 - Bing AI 2 - Canva, then 3 - Nightcafe because it includes mazes and is situated at the correct SCJ angle. 4 - Dream by Wombo - nice market…no maze.

Prompt 7 - Make a maze of a slice of an orange in color

This prompt was poorly written by me I guess. Lots of garbage output IMO. Our finalists are ToolBaz, Dall-E, Nightcafe and Bing AI. These made the cut because they include mazes and oranges. I do not feel strongly about the order after Bing AI…how about ToolBaz, Dall-E, Nightcafe ?

Prompt 8 - Make a maze integrated on top of a photograph of a king sitting on his throne looking cantankerous beside his beautiful queen

Using the term photo made this interesting. I also got a lot of horror blurred faces generated. Overall I was not impressed with these results. StarryAI, Dream by Wombo,Dall-E and Bing AI are our finalists. 1 - StarryAI because he looks like a deformed James Bond character, 2 - Dream by Wombo because the king looks like the most interesting man in the world, 3 - Bing AI for that king and 4 - Dall-E for including the queen as instructed.

Prompt 9 - Make a solvable maze that is very large and very difficult to solve because it is so complex

This was interesting from an output perspective. If an AI could shine with making makes this is where it would have come out. I wanted to see a large maze generated and a few actually delivered. We have Canva, Nightcafe and StarryAI, Dream by Wombo. For me 1 - Nightcafe - this fantasy labyrinth around a castle would have been a great output for #4 above (which Nightcafe won anyway !), 2 - StarryAI - if you could read the labels this would be better, but the maze is doable, and 3 - Canva because it looks like a complicated maze, but if you really attempt it…uh not so much. 4 - Dream by Wombo - just a solid 3D maze.

Prompt 10 - Make a 3d render of a red and blue glossy cube maze

This was a prompt that many AI generators exceled at creating interesting images for. For most I generated 4 images and had trouble picking the best among them..so I bent the rules here and picked a top 5. StarryAI, Nightcafe, ToolBaz, Canva, and Dall-E from top to bottom left to right. Ranking is tough, but I’ll go 1 - Dall-E for so many nice options, 2 - ToolBaz - it’s a small maze but make it bigger and I want to solve it, 3 - Canva - a crisper version of 4 - StarryAI, and 5 - Nightcafe.

That does it. Let’s tabulate our results and see how everyone did. I will go with a 4,3,2,1 scoring system, with half points for 5th place on the last prompt…time to make a spreadsheet. And here it is:

Congratulations to StarryAI, Bing AI, Nightcafe and Stable Diffusion with the highest scores. This fits with what I found. And it seems like we have 3 sections in the results: The 4 best AI generators, the 4 middle class, and the 4 poor performers. I would say these findings are only for making maze art. Each generator can do wonderful things with the right prompt.

To remind you about the top 4:

StarryAI : You must sign in to use the site and you receive 5 credits per day for free. Each generation will generate 4 images and cost a credit (If you want 8 variations that will cost 3 credits). There are 6 styles to choose from, no style, Fantasy, Portrait, 3D Art, Wooltize and Anime. You may also upload a photo to give the AI “something to start with”.

Bing AI Image Generator: Sign in using your Microsoft account. Using the site is free to create images. Powered by DallE, but when I used it seemed to do a better job than DallE did with the prompts (I do not know why).

Nightcafe : You get 5 free daily credits to create with. Each prompt, or generation will generate only 1 image. It features 4 different style algorithms to use.

Stable Diffusion : You can try the site without signing in. Each prompt, or generation will generate 4 image options. There appears to be no limit to use for free.

I would also mention that 5th place Dream by Wombo only generates 1 image at a time vs. most other sites generating 4. I thought it was consistently strong despite being at a disadvantage. How did you score the sites?

The final blog post in the series is next:

What I learned using AI to make maze art