‘Add Spores and Mushrooms’ is a personal project exploring the function of AI in Photoshop by combining typography and generative fill. It’s part of a wider focus developing my creative practice considering both typographic treatments as well as understanding the parameters of artificial intelligence when used in graphic design.
Having decided on the route I wanted to take, I chose the theme of spores and mushrooms as I wanted something natural and organic, a subject that felt in contrast to the technology at play.
I gave myself strict boundaries to work within, I chose not to alter any colours, add filters, or make any post-‘fill’ modifications. And most importantly, I only allowed myself to instruct the generative fill prompt with the single phrase ‘add spores and mushrooms’.
I systematically worked through each of the 26 uppercase characters starting with the letter ‘A’. I carefully considered the choice of background colour, knowing it would influence the generative fill outcome. I’d then select small sections with the lasso tool, asking the generative fill to ‘add spores and mushrooms’, and bit by bit, slowly build up the image.*
From the three generative fill results that appeared each time, I’d choose one and delete the other two, moving swiftly to the next selection.
I explored how different outcomes were generated depending on the selection I made, whether it was an area away from anything already in place, or a section that overlapped elements currently there. As I slowly filled the letterform, elements would start to meld and interweave, shapeshifting and becoming something slightly different to what had previously existed. I imagined the letterform being like a glass structure, a vessel that was slowly filling, comparable to a test tube.
In the early part, there were occasions when my impatience crept in due to the wait time for the AI to do its work. Keen to get results fast and see what would appear, I’d select either far too large an area or lots of small areas. In both instances, the results were never satisfying. They wouldn’t fit the overall feel and would compromise the emerging image, sometimes throwing it in an entirely different direction. This dictated how the process should go; select small areas, and slowly build the letterform.
As I worked through each letter, my decision-making got faster and would often take less time than the AI itself.
From sparse beginnings where the initial generative fill didn’t look very engaging, I’d repeat the process, slowly building the letterform. There was often a turning point where the image became more visually appealing, providing a clear style and direction. I felt it was important not to let the elements sit too tightly and to encourage the objects to breathe and remain connected.
Some frustrations came when an object was a repetition of imagery already used, or an outcome was out of focus and blurry. I would have to delete the layer and start over.
I started recording the time each generative fill took. Every time I hit ‘generate’, it would take between 10 to 15 seconds to give me results, sometimes a little longer. I recorded the total time it took me to create the ‘K’, the 11th letter in the alphabet and the 11th file I’d worked on. The initial generative fill was actually very attractive, but the results started to go in an entirely different direction, one that I wasn’t keen on. It felt challenging, almost as if AI was bored of my repetitive requests of ‘Add Spores and Mushrooms’. I felt I had to tussle with the AI to bring it back round to something I was happy with. 40 minutes later and with 44 generative fill layers, I concluded the ‘K’.
Another challenging moment was when I lost the completed letter ‘I’. This was particularly frustrating as I’d been extremely happy with the finished results. The incident occurred in between backups; there was no file to recover. It was gone. It made me realise how fragile the process was, the AI providing entirely unique outcomes every time it was prompted.
The new ‘I’ began its life dark and murky, almost as if it could feel my frustrations. It took a lot of work to get it to a point I was happy with.
The constraints I had set myself, in that I only allowed the prompt ‘add spores and mushrooms’, meant I couldn’t utilise the tool in a way that would be to my advantage and the way it has been designed to be used. On numerous occasions, it would have been helpful if I could’ve given specific instructions such as a change of colour, to remove or replace an object, or give a more accurate description of what I wanted to generate.
Occasionally the generative fill would throw up something entirely in the wrong ballpark, often a picture book style illustration rather than a photograph consistent with the rest of the image, or a manmade object that was entirely unrelated. These are known as AI hallucinations, when a generative AI model produces inaccurate information believing it is correct. Before deleting them, I screenshot a few for reference. They included an orange plastic bucket, a woven cane basket, and what looked like a white plastic diaphragm with an elastic strap.
Some lovely moments came when nearing the completion of a letter; the final stages would involve filling in the smallest of details to help balance the letterform. The generative fill would throw in an object with a slightly different style or colour that would nestle into place, interacting with its neighbours. Sometimes however, when the letterform was full, I’d make the tiniest tweaks to try and improve it, but the AI often didn’t bring much more to the image, almost replicating what was already there. This was when it felt an appropriate time to consider the image complete.
As my understanding of how to use the tool expanded, I considered the thought that the AI was learning from me all the time I’d been creating these images.
I completed the suite of 26 letters over three days. I developed a process whereby I was partnering with the AI, curating the generative fill results it gave me while it provided me with content I couldn’t have achieved on my own. It felt like an entirely equal collaboration, a joint venture with a satisfying outcome.
I’ve already started to explore the results further. The obvious step being to form words and sentences from the letterforms, as well as to run the process with different fill prompts to see what other results emerge. However, I’ve also been looking at reducing the letterforms to a silhouette, forming something that might lean into becoming a font, stripping everything right back from the elaborate AI visuals leaving very little trace of the journey the letterform had been on.
Notes on my process
* I developed a system for each of the 26 uppercase letters. I’d fill the background with either a flood of colour or use the brush tool to loosely paint the area, leaving very little uncoloured. Occasionally I combined both methods allowing a little variation in tone or colour to creep in. The background colour would influence the outcome of the first generative fill, each time throwing up varying results, it was interesting to see how the choice affected this.
For the letterform, I chose a sans-serif font in extra bold. I positioned it on a hidden layer placing guides to roughly indicate where the letter would sit. The positioned letter itself would need to be hidden otherwise, the colour of the type would influence the generative fill outcome.
I then selected small sections with the lasso tool and asked the AI to ‘fill’ the area using my unchangeable prompt ‘add spores and mushrooms’. The first selection couldn’t be of significant size. Through trial and error, I found selecting more than 20% of the letter would throw up the wrong results, or at least results I wasn’t happy with.
The Photoshop generative fill offers three outcomes. I’d learned from the ‘tips’ that popped up while waiting for the results that it would help with the overall file size to delete unused generative fill options, so this became part of the process. Choose one and delete the other two. I’d choose mindfully as each choice would lead the way visually, dictating the colour and style of spore or mushroom on the next set of results. At the same time, I made sure I wasn’t too precious about the decision, I’d quickly commit to a direction and not overthink it.
The generative fill layers averaged about 35 per file, ranging from as little as 18 layers to a significant 58 layers!
コメント