Rosalia had never been a daydreamer. Nor had her dreams at night meant much to her or stuck with her through the days. She was fine with this, blessed with an innate talent to find fulfillment in the pragmatic. What life threw her way, she enjoyed tackling, even when it felt like life was throwing an abundance of pipe wrenches at her ankles. She had no time for fantasy, nor inclination to lament its absence. She did, however, appreciate beauty when she saw it, and was, like anyone she knew, drawn to it and inspired by it. She had not considered this tendency a weakness until after it led her to slip on the black ring (or had it led the ring to her?) The voices it cast into her head threw her life into immediate disarray. They offered her magical solutions to practical problems she thought she'd already solved, forcing her to second guess her original choices. The solidity of her existence dissolved into a roiling ocean of doubt, against which she struggled mightily just to keep breathing. After less than two days, she threw the ring away, without even turning to see where it landed. But later that evening, it was on her finger again. She disposed of it at least a dozen times more, literally flushed it down the toilet (more than once), yet it always reappeared. Finally, she just stared it down, glaring at her hand in silence until the ring faded away on its own. She considered this a victory, until she noticed the ring's parting gift, at first in the bathroom mirror: numerous colors in her hair that she hadn't put there. Then the colors were on the mirror, and around it, and spreading onto and staining everything in sight. A pointless, cumbersome power no one had asked for, or likely ever would have. But she knew the game. The ring wanted her to find it again. Only now was she compelled to care enough to ask why.
Rosalia 48C0C8
Rosalia 48C0C8 (LVL00)
Rosalia 48C0C8 (LVL1~)
Rosalia 48C0C8 (LVL2~)
Rosalia 48C0C8 (LVL3~)
About My Process
This set of images represents the final plateau of my increasing frustration with DALL-E 2. For reasons that remain mysterious (by this point, I've come to assume that, like the "black box" functionality of generative AI models themselves, these reasons will simply remain unknowable), after around 3 solid months of constantly messing with it and being repeatedly surprised at and delighted with (for the most part) the results I was able to produce, DALL-E 2 just… completely lost the thread.
Whereas my initial set of experiments (resulting in 50 image sets I was pleased with vs. only around 5 that I considered failures) felt like nothing less than a collaborative/creative partnership with the machine "intelligence," filled with the almost spooky sense that I was actually trading ideas back and forth with it, around Halloween 2022 it all went down the tubes, making my original workflow instantly, completely unworkable, as it depended so heavily on the AI spitting out results that were not complete and utter bullshit. That 90% success rate was suddenly a thing of the past. Within the next month, I had 70 new images in progress, but only around 5 of them (a mere 7% of the total attempted) had come anywhere close to a satisfactory finished state… It would take another six-plus months of detours, including plenty of experimentation with other AI engines, to cobble together a new workflow and once again start to produce results I was in any way happy with, and I've had to accept that the majority of those failed intermediate image sets may never be salvageable. Many of my experiments have involved repeated attempts at just that, but so far it has often proven too difficult to rescue these girls from their shit-image prisons, and in those cases, I've just pivoted away and worked on a different and/or new face, the same keeping-myself-productive technique that created those 70 new images in the first place…
Rosalia here is a rarity, in that she was started in DALL-E 2 as its enshittification was occurring, but as I was working on her during the most intense period of initial frustration, before I had nailed down other workable solutions (and months before Adobe would integrate the Firefly-powered "generative fill" directly into Photoshop), she basically got transported into a stock image background (of an apparently real-life Brooklyn brownstone), given stock image legs and feet, and more or less bitterly Photoshopped until I felt a sufficient amount of transformation had taken place.
For the next year, from this point forward, I hoped that the eventual release of DALL-E 3 would come to my rescue, but instead, two quite different outcomes developed. On my end, I spent so long experimenting with other, multiple engines (during this phase, primarily Midjourney and mage.space, powered by Stable Diffusion / SDXL) that I began to feel more confident and proficient with them, eventually able to produce far better / more detailed images than DALL-E 2 had ever been capable of approaching. On OpenAI's end, ChatGPT became the talk of the industry and their flagship product, and clearly the project behind which they threw the majority of their resources.
By the time DALL-E 3 finally appeared, it couldn't help but instantly pale in comparison to its competitors, all of which had released multiple new updates in the meanwhile. DALL-E 3 did show a somewhat improved understanding of complex prompts, and was slightly better at reproducing coherent text within generated images, but its image outputs tended toward a weirdly specific, tell-tale, overly-illustrative feel that rarely appealed to my personal tastes. In direct comparison with images produced in other engines using the same prompts, it was only occasionally going to prevail. OpenAI also made it impossible for me to return to my original workflow with their new engine, as they inexplicably chose to abandon the clean, elegant infinite canvas interface of DALL-E 2's "labs" site. At first, in both the free and paid versions of DALL-E 3 (the former implemented as Bing.com's image creator, the latter paywalled behind and integrated with a ChatGPT subscription, natch), there was zero additional functionality available aside from prompting. No inpainting, no outpainting, no editing or expanding existing images. Back to only "one click and done." I am pretty sure the paid version now allows for at least inpainting, but I stopped paying for ChatGPT after a month or two of highly frustrating experimentation. Just let me directly edit the damned images, not be forced to depend on a constantly lying large language model to rewrite my text input as a text prompt (why, OpenAI, WHY?) to submit to DALL-E 3 for an entirely new image. Just another annoying instance of the makers of this tech positioning it as a tool to completely replace artists, rather than as another tool for artists to use.
