In this blog post, we wanted to walk through our process for creating a custom commissioned AI art. If you’re interested in inquiring about this service, first have a read through this post, then check out the custom AI Art Commissions section of the main site.
Why Custom AI Art?
One of our potential customers was a couple that had done extensive worldwide travel together. Throughout the years, they had collected a variety of statues, objects and other memorabilia which they had placed above a fireplace mantle, but they had never found the right picture or artwork to go along with the travel-themed space in their house. They were interested in creating a large piece of landscape artwork chronicling their past travels as a couple. Their goal was to piece together a collage of memorable scenes in way that blended seamlessly and integrated the photos, but with a painted artistic style.
Step 1: Gathering Source Materials
To help us get a feel for the scenes, our client provided a variety of photographs from each of their many journeys – these included destinations like …
- The Great Wall of China
- Skogafoss waterfall in Iceland
- Mount Kilimanjaro, Kenya
- The Hagia Sophia in Istanbul
- The Parthenon in Greece
- Machu Picchu in Peru
- Kiev, Ukraine
While individually each of these images might be something an AI artist could paint, combining them together in a coherent fashion had the potential to create challenges.
Step 2: First Attempts to Create AI Art
Because of the unique mash-up of these destinations, we felt it was very unlikely that an AI artist would be able to successfully create a work of art entirely from scratch, which distinctly features all of these elements. Or at least not in a fashion that humans would like!
In early trials, we found that some elements were either too obscure (Skogafoss) to be correctly captured, and at the same time certain elements had a tendency to merge – for example the AI artist might create a great wall of China with minaret-like structures on it, rather than distinctly creating the Great Wall connecting to a distinct Hagia Sophia Mosque. Similarly, it had a tendency to make misty waterfalls all over the image.
Here are two examples of early drafts, which we didn’t like much … nevermind that we hadn’t settled on painted style attributes just yet (these were too photorealistic for our client)…
While too photorealistic, elements of this image also tended to merge parts of the Great Wall of China into a Parthenon and Hagia Sophia simultaneously.
The foggy effect here was interesting, but it evoked the look of waterfalls throughout the image, and led to less distinct architecture.
Step 3: Giving the AI Artist More Specific Direction
To resolve these issues issue, we mocked up a very crude layout of the images in photoshop in a fashion that might allow for more natural transitions between images – while also ensuring that each element was accounted for in its own distinct area. For example, placing the mountain of Machu Picchu in front of Kilimanjaro made sense, and it also made sense to place the sky of the Parthenon along Kilimanjaro’s skyline. Because we weren’t sure how the final output would materialize – we created a couple layout versions before narrowing it down to a favorite with our client.
By crudely positioning key elements as a starting point, it would allow us to give very specific direction to the AI system for positioning key travel elements.
Step 4: Running some Trials
Once this template was in place, we let the AI artist go to work. We experimented with several styles, levels of initial noise and various other prompts to see which layouts produced the most interesting outputs. In some cases we found that while the image looked great, certain colors were a bit off. For example, the artist had a tendency to make the Parthenon a golden hue, instead of retaining a white or grey tone. Similarly, the green mountains of Machu Picchu had a tendency to take a more rocky look, or incorporate more Great Wall elements than we or the clients wanted.
Image is looking good, but the Parthenon is both too golden and missing pieces, while Machu Picchu has acquired attributes of the Great Wall. Similarly Skogafoss has disappeared from the lower left corner.
Step 5 - Photoshop Reworking
To resolve this issue, we would take some of the more promising outputs from the AI, and manually adjust parts of them in Photoshop to adjust hue, contrast or other elements, before re-submitting to the AI to continue its work.
When working with a human artist, we often take for granted the ease with which you can say “Make that object - yes that one there - a bit more white.” With an AI, sometimes you need to use a different strategy to get specific, in particular by providing it an example in image form to work from, as human language may not be precise enough when dealing with thousands of pixels.
After several training sessions the work was finalized and signed off by the client and sent to production.
The final image (low resolution)
The full-size image above is substantially lower resolution than what we actually send to the printers (which was 12,000+ pixels wide and nearly 140 MB). For a bit of further context - below are some zoomed-in crops of the actual scene, showing some of the finer details and nuance in the actual final image.
Close-up details of foliage
Close-up details of Kiev church
Close-up details of the Parthenon, including detailed brush strokes
Step 6: Enjoy!
Our clients completed artwork arrived at their home 9 business days after signoff. The piece was mounted above a fireplace using the supplied mounting hardware, making a fine addition to the couple’s adventure travel-themed mantelpiece.
If you found this interesting – you can reach out to discuss your own custom art commission.
For pre-existing original AI Art, check out the Artaygo home page.