Updated: Jul 27
While many Pixelmind series offer new, distinct styles in which any subject can be explored, I’ve always found myself attracted to more specific forms of co-creation. My series “Doorways to Anywhere”, launched on February 11th, 2022 but by that point had been in development for more than three months. The original idea was inspired by our early experiments with Guided Diffusion. Simply typing in “The Doorway to Universal Truth and Galactic Meaning, abstract, trending on artstation” in November of 2021 yielded jaw dropping results and the seed of the series was planted as I explored variations on the phrase.
But while the results were stunning, one thing quickly became clear - The placement of the door itself, whether it was open or closed and many other critical details were all over the place. The same prompting phrase (“The doorway to geothermal hot springs and dwimmercraft” for the images above) would generate a great door and then the next create a door standing alone in a pool of geothermal water! Variety is good but it was too much. I needed a starting point, and more importantly I wanted to see what was beyond the doorway, not just look at the door itself.
I needed a way to tie the series together and give co-creators the best possible chance of creating something amazing, every time.
To accomplish this, I started looking around my favorite copyright-free image sources, namely Unsplash.com. There, I found a wide variety of images which would fit the bill, providing structure that would create some uniformity in how each creation would develop, while leaving plenty of room for an individual's creativity to shine through.
I tried starting from many photos in my hunt for structure but eventually settled on an image by Rowan Freeman, taken from inside an arched stone doorway looking out at the Italian countryside.
The image had much to recommend it. The doorway was apparent in silhouette with strong, colorful details and lots of complexity that would be interpreted by the AI in different ways each time, while keeping the overall “thing-ness” of the piece consistent. But the most important factor was the lighting on the rough-hewn stone floor. When working from a starting image, light and shadow are arguably more important than any other element of an image.
When tuned with the appropriate number of skip_timesteps (I eventually settled on 35), the AI tends to adapt the rough texture and lighting to any given prompt, adding significant perceived realism to the resulting creation.
So I had my starting point and the series began to come together, but I found myself frustrated by the not-infrequent occurrence of closed doors, and even when the doorways were open I found the tall, narrow image to limit what could be seen on the other side. I wanted to see more.
After much experimentation, I did the easy thing that I actually expected to be wrong. I squished the image in series settings. While the starting image itself was about 1050 pixels tall by 750 pixels wide, I simply increased the width of the images that would be produced to 1050. So I was using a starting image that was taller than it was wide, but I was using it to produce images which were perfect squares.
This approach kept all the things I really liked about the starting image but by widening the door, I was able to see more on the other side and achieve my broader goals. This completely eliminated any closed-door generations and emphasized the truly unique element of each co-creation, which is not the door itself but the location it takes you to.
Now I had a series that would allow you to travel to Dracula’s castle, the surface of an alien planet, a Japanese tea garden in the 1800s, balmy Pinecrest Florida, the Taj Mahal or really any place. That’s the magic of creation with AI. You often don’t know exactly where you’re going, or even how to get there, but the combination of curiosity, technology and perseverance can take you anywhere you can imagine.
See the entire series HERE