A few months ago I generated forty variations of the same idea in about twenty minutes.
It was an old Miami storefront, the kind of subject I keep coming back to, and I typed a description into an AI image tool mostly out of curiosity. Faded paint. Hand-lettered signage. Late afternoon light. Within seconds I had a grid of images, and within a few more minutes I had forty of them, each one technically different, each one competently rendered, each one entirely forgettable.
None of them looked like mine.
They looked like anyone’s. That was the uncomfortable part. If I had shown that grid to another artist working in a similar space, half of them could have plausibly claimed any single image in it. There was nothing in the file that pointed back to me specifically, and I had spent zero effort getting there. The tool had done all the work, and the work it did was, by design, generic enough to belong to no one in particular.
This is where a lot of artists get stuck with AI-generated imagery, and I don’t think the stuck point gets discussed honestly enough. The conversation online tends to split into two unhelpful camps. One camp treats the prompt as the finish line, as though describing an image well is the same thing as making it. The other camp dismisses AI-assisted work entirely, as though using the tool automatically disqualifies whatever comes out of it from being real art. Neither position matches what I’ve actually found sitting with these files for the last couple of years.
The honest answer is closer to this: an AI-generated image is a starting point, not a finished decision. It is closer to a rough block-in than a photograph. What happens after generation is where authorship actually lives, and that stage gets skipped constantly, mostly because the image already looks so complete the moment it appears on screen.
I want to walk through the process I actually use now, because I think the shortcut version, “just add your own style,” is about as useful as telling someone to “just find their voice.” It’s true. It’s also not a process. Here’s the process underneath it.
Treat the Generated Image as Reference, Not Result
The first shift that changed how I work with these tools was refusing to treat the output as a finished file, even when it looked finished. I started opening every generated image the same way I’d open a reference photograph pulled for a collage: as raw material to be taken apart, not a picture to be lightly touched up and exported.
This sounds like a small distinction, but it changes what you do next. If an image is “basically done,” you make small adjustments, a contrast tweak here, a crop there, and you stop as soon as it looks clean. If an image is “raw material,” you go looking for what to keep, what to discard, and what needs to be rebuilt from scratch. The second posture produces something. The first posture produces a slightly adjusted version of what the model already gave every other person who typed a similar prompt.
Find and Remove the Tells
Every AI image generation model has habits, and those habits are the fastest way to spot work that never left the tool. Certain kinds of symmetry. A particular way skin or fabric texture gets smoothed. Backgrounds that are technically coherent but emotionally empty, as though the space behind the subject was generated to satisfy a prompt rather than to support a mood. Lighting that is dramatic in a generic, stock-photo way rather than specific to the actual subject in front of it.
None of these are flaws exactly. They’re just fingerprints of the tool rather than fingerprints of the artist. The first real editing pass I do on any generated image is a search for those tells, the same way I’d look for compositional weaknesses in a photograph before I start refining it. Where is the eye landing because the model defaulted there, rather than because I decided it should. What reads as impressive but says nothing specific about this particular image.
Removing those tells usually means compositing in your own texture library, adjusting the grain and tonal relationships by hand, breaking up whatever symmetry the model reached for by default, and often replacing entire sections of the background with something built rather than generated. This is slower than it sounds, and it’s supposed to be. The slowness is where the file stops belonging to the tool.
Apply the Decisions That Are Actually Yours
I’ve written before about the two or three load-bearing decisions that make a body of work recognizable, things like a consistent color temperature, a specific grain treatment, a habit of restraint in how much detail gets left in a frame. Those same decisions are what separate an AI image that’s genuinely yours from one that technically came out of your account.
The test I use is simple. Strip away the subject matter and ask what’s left. If the only thing connecting a generated image to the rest of your work is that you typed the prompt, nothing is actually left. If the color temperature, the tonal restraint, the grain, and the compositional habits you already protect in your other work are all present and doing the same job they do everywhere else, the image has been pulled into your visual language rather than sitting next to it as an import.
This is also where a texture library and an established editing process earn their keep. An artist who has spent years building a consistent way of handling shadows, color grading, and grain has something the model doesn’t: a set of decisions that stay constant regardless of the source material. Running a generated image through that same process is what actually converts it from output into work.
Composite, Don’t Just Adjust
The single biggest shift in my own process was accepting that a generated image is often better used as one layer among several rather than as the whole file. A background from one generation. A subject pulled from another. A texture scanned from an actual physical surface, layered in by hand. Typography added the same way it would be added to any other piece. None of this requires the final image to look like a collage in the literal sense. It just means the file stops being a single AI output and starts being a composited piece the way any other mixed media work is composited, with the generated material treated as one ingredient rather than the entire meal.
This is uncomfortable for artists who want the process to be fast, because it isn’t. Compositing takes real time, the same amount of time it would take with any other mixed media piece. But that time is exactly what the shortcut version skips, and it’s exactly what makes the difference between an image that says “I typed a good prompt” and one that says “I made this.”
Ask the Harder Question Before You Call It Finished
Before I let a generated-and-refined image into a collection now, I ask one question that has nothing to do with how polished it looks. Could someone with the same prompt, using the same model, have produced this exact file? If the honest answer is yes, the piece isn’t finished, no matter how clean the color grading looks. If the honest answer is no, if the file carries decisions a different person with the same starting point would not have made, then the generation did its job as a starting point and the rest of the work is actually mine.
That question has quietly kept a lot of technically impressive images out of my own portfolio, and I think it’s the most useful filter I’ve found. It doesn’t ask whether AI was involved. It asks whether anything happened after AI was involved.
Where This Leaves an Artist
None of this is an argument against using these tools. I use them regularly, the same way I use a scanner, a camera, or a stock texture pack, as a source of raw material that still has to be shaped into something before it means anything. The mistake isn’t reaching for the tool. The mistake is stopping the moment the tool hands something back that already looks presentable.
An image that looks finished and an image that is actually yours are not always the same file. The gap between them is where the real work happens, and it’s usually invisible to anyone looking at the result, which is exactly why it’s so easy to skip. Nobody can see the compositing pass, the texture library, the hours spent removing a tell that only another artist working in the same tool would recognize. They just see whether the finished piece feels like it belongs to someone in particular, or whether it could have come from anyone who typed roughly the same words into the same box.
Generation gets you a starting point faster than it’s ever been possible to get one. It has never made the rest of the work optional.
About the Author
Orlando Monteagudo combines analytical thinking with mixed media experimentation, Photoshop workflows, AI-assisted creativity, and practical digital refinement systems designed to help artists create more cohesive, polished, and sustainable creative work.
Keywords: making AI art your own, AI image editing, AI art refinement, personalizing AI generated images, AI art workflow, Photoshop workflow for artists, mixed media digital workflow, compositing AI images, AI art authorship, digital art refinement, visual identity in AI art