Welcome to the 3rd in a series of posts where I will test AI image generators and see how they handle making maze art. I will be asking 10 prompts and seeing what gets generated. My goal is to evaluate different AI image sites against each other to see how they perform. If you need to catch-up here is the series so far:
An exploration into Al Image Maze Generation
Making mazes with AI: DallE2 (Updated to compare 2023 to 2026)
Update from 2026:
When I did this series in Spring of 2023 I found that none of the text to image generators were able to generate actual mazes, including this one. They were able to make images that looked maze-like, but none were solvable puzzles. The 10 prompts below are a record of what was possible then, but each generation of AI pushes the limit further on what is possible. Feel free to try some of the prompts below to compare how far the development has come. All of these were made using text to image generation and not coding.
I also did an update using Stable Diffusion in December 2024 (18 months later) of all of the prompts to compare. Those images are included here as well.
Making Maze Art with stable diffusion
2023: Today I test out Stable Diffusion. You can access the Stable Diffusion website here. You can try the site without signing in. Each prompt, or generation will generate 4 image options. For this exercise I chose the one closest to what I had asked, or the most interesting. All prompts I use are the same for each site !
2024: I did quick generations and generated only 1 image per prompt (last time I picked the best of 4). I generated images for all 10 prompts.
2026: The site asked me to sign in, and gave me 10 credits to use. I used the Base model and generated 1 image per prompt. There are models marked Plus, Pro, and Ultra which are paid versions. There are 6 styles to choose from and I used AUTO for all since the prompt often includes a style. I generated images for the first 3 prompts, shown below.
Prompt 1 - Make a medium difficulty maze of the Eiffel Tower in black and white with arrows at the start and finish
2023: That is the Eiffel Tower. It is black and white. I see no arrows, and the maze is not really a maze, but more like placed blocks. Overall it is an interesting image though.
2024: Our new prompt did a better job executing what was asked for. I like the contrast of the white on a grey background.
2026: I notice how similar the 2026 version looks when compared to the 2024 version.
Verdict: Most of the improvement here came between 2023 and 2024. The 2026 image does look the most like a traditional maze, in black and white.
Prompt 2 - Draw a medium difficulty large maze of the Empire State Building with the start and goal embedded in the structure
2023: The maze is both inside and outside the building that looks like a short ESB. I like how the building needle is off the top of the page. Interesting !
2024: I think the use of color in the new version is excellent. If you look at the top of the ESB you can see START on the building. There is even an attempt at a maze. The city stretching out behind the ESB is cool. 2024 wins easily !
2026: I think the generation is very interesting. Instead of imbedding the start and goal into the maze, the entire building was imbedded into a maze. The maze appears to be very well done except for the missing start and goal.
Verdict: Close. The Empire State Building looks excellent in the 2026 version.
Prompt 3 - Draw a difficult maze of the White House pixel art style
2023: That is a picture of the White House in a pixel style mixed with a bit of a gridded style for the lawn. Weird, but cool.
2024: 2024 lost the antennas that 2023 had, so that is an improvement. The marker style from 2023 was cool, but there was no maze…and barely one in 2024. The White House looks better in 2024 though, so 2024 wins again.
2026: That is technically a maze AT the White House, not OF the White House. I think it is an interesting solution. It is pixel art, and the building looks correct.
Verdict: Much improvement over time. And this might technically be out first actual solvable maze, despite being small.
Prompt 4 - Draw a difficult maze that looks like a drawing of a famous building in sketch style
2023: Looks like a museum or parliament or a stately residence. Oh, and there’s a small maze out front. Cool.
2024: A sketch style with a nice walled maze. Pick the style you prefer.
2026: Appears to be a sketch of a church façade from Italy. I do not recognize the building. There is no maze but I like the illustration.
Verdict: When you are less specific with your prompt it is interesting to see what the AI generator comes up with. I enjoy that each of these is so different. I do not see an improvement from a maze perspective.
Prompt 5 - Draw a maze in the style of doyoumaze.com of a skyscraper in NYC
2023: Not my style, but I like how this looks despite the obvious miss with the maze. Looks like a bunch of buildings squished together.
2024: 2023 was a squished building blob. 2024 did a nice ESB maze like image. Easy winner.
2026: Another ESB illustration that is very good, this one surrounded by an almost maze. Close
Verdict: None used my style, maybe a good thing, with all 3 images having different looks. The best illustration and closest to what was asked for was from 2026.
Prompt 6 - Draw a maze in the style of Sean C Jackson of a scene from a large outdoor market
2023: Not an outdoor market when I look at it - more of a labyrinth. This is my favorite image generated so far. I would use this layout to make a maze of a city.
2024: Despite being interesting this is a fail.
2026: I am a fan of the use of color. This is a market with a maze in the middle, although it looks good, the maze is not solvable.
Verdict: 2026 is the closest to the prompt yet, and the great colors and market make the image interesting.
Prompt 7 - Make a maze of a slice of an orange in color
2023: They can’t all be winners.
2024: Double fail, but 2024 comes in second to last by a hair.
2026: Honestly I had given up on this prompt as just being poor to ask based on having seem what was generated by many different generators over time…..and yet. I think the 2026 version is technically correct. It is what I asked for, and the maze is solvable, but with 3 gaps in the outer wall. I think with some prompts this could be adjusted and actually work…but I only have 10 credits to use. Feel free to download the image and try it yourself.
Verdict: So close to what I would expect in 2026. Maybe this will work soon.
Prompt 8 - Make a maze integrated on top of a photograph of a king sitting on his throne looking cantankerous beside his beautiful queen
2023: Remember I picked the best one out of 4 generated. A maze with cursive letters on it, kind of, and a double faced king with one crown next to his red faceless heavily shadowed queen. Perfect.
2024: Versus what 2023 spit out, 2024 wins this one by a huge margin. It’s a double throne chair and I like the maze being held together on the lap. But did you notice they are almost wearing one big dress ? You have to look close in the shadows to see a change in the pattern.
2026: Who needs a queen. Good looking king. We have one opening on the maze, but not 2 or a real goal inside. This would be good if there was a queen around.
Verdict: Sometimes parts of the prompt get ignored. 2026 and 2024 each have their strengths and weaknesses.
Prompt 9 - Make a solvable maze that is very large and very difficult to solve because it is so complex
2023: Well. We had a nice start didn’t we ? I like the contrast of the colors and the use of 3D without prompting.
2024: An interesting 5 point perspective wooden box maze. I like how the illustration is creative.
2026: This looks like something someone might actually use in a blog post or as a thumbnail. I like the overhead perspective and the internal shadows. There are 2 outer wall breaks in the maze ! But they are not connected. So close.
Verdict: With some help I think the 2026 version could be fixed. Much improvement from 2023.
Prompt 10 - Make a 3d render of a red and blue glossy cube maze
2023: I decided to show all 4 once again. All are so different. None are really great but together they are fun to look at.
2024: 2024 only failed once, but 2023 failed 4 times, so there’s that !
2026: That might be a maze inside there. It is closer to the prompt than all of the other options.
Verdict: We saw improvement over time, and maybe our first real maze ( but only 1 out of 10).
I am out of credits.
Summary and Conclusion:
How did Stable Diffusion do ? I Think it could handle famous buildings but struggled making a real maze. Overall some interesting images. We saw improvement from 2023 to 2024 and even more to 2026. I think many of these generations could be used as seeds to generate even better versions - all of these were one pass.
Coming Next: Craiyon
