Type “oil painting of a lighthouse in a storm, thick impasto brushwork” into a text box, wait about eight seconds, and four finished images come back. Two are genuinely striking. One has a lighthouse with two roofs. The fourth looks like a stock photo someone dragged through a paint filter.
That gap between striking and unusable is where AI painting lives right now. The technology is remarkable and stubbornly uneven, and the people getting good results are the ones who respect both halves of that sentence.
What the phrase actually covers
AI painting is a bucket term. It takes in the free phone app that turns your selfie into a watercolour, and it also takes in a three-day pipeline where a working illustrator builds one book cover using depth maps, hand-painted corrections, and a reference folder full of 19th-century landscape plates. Calling both of those the same thing is a bit like lumping together a graphite sketch and a fresco.
Underneath, most of the software runs on diffusion models. The short version: the model studies hundreds of millions of image-and-caption pairs and learns to reverse a process in which noise gets gradually added to a picture. To generate something new, it starts with pure static and strips the noise away step by step, nudged along by your prompt, which a text encoder has already converted into numbers.
It is not one tool, and the tool matters less than you would think
Ask five artists what they paint with and you will get five different answers, often from the same person in the same week. The choice matters, but a strong prompt in a mid-tier tool beats a lazy prompt in the best one every time.
The tools worth your time
- Midjourney. Still the strongest default aesthetic. You get fewer knobs to turn, which is either a relief or a frustration depending on your temperament.
- Stable Diffusion, SDXL, and Flux. Open weights you can run on your own machine. Slower to set up, far more controllable once you are there.
- DALL·E 3. Best in class at following long, literal, multi-clause instructions, which makes it handy for diagrams and awkward specifics.
- Adobe Firefly and Photoshop’s Generative Fill. The advantage is location: you are painting inside layers you already have, not exporting back and forth.
- Ideogram, Leonardo, Krea. Each has a niche. Ideogram handles lettering better than most, Leonardo leans toward character work, Krea offers a real-time canvas that responds as you sketch.
If you want a deeper read on why one of these has become the default for so many illustrators, this breakdown of what makes Midjourney the artist’s choice for AI images is worth twenty minutes of your evening.
Why so much AI painting ends up looking identical
Scroll through any public gallery of generated art and a pattern shows up fast. Soft rim lighting. A figure standing centre-frame, back to camera. Colours that sit in a narrow teal-and-orange band. Everyone is typing roughly the same prompt because everyone learned from the same trending posts.
The words “cinematic lighting, hyper-detailed, 8k, trending on ArtStation” were useful for about six months in 2023. Now they flatten everything into the same glossy sheen. If your goal is work that reads as yours, the fix is usually subtraction. Describe the light source, not the mood. Name the medium, not the quality. Feed in three reference images you actually like rather than ten adjectives you copied.
Getting an image you would hang on a wall
Write in sentences, in this order
Subject first, then action, then setting, then light, then medium, then the technical camera or canvas detail. “An old woman mending a fishing net on a harbour wall, low winter sun behind her, oil on linen, visible brush texture.” That single sentence outperforms a comma-separated list of forty keywords, and it is much easier to edit when something comes out wrong.
Then take control of the composition
Text prompts alone give you suggestions, not instructions. The moment you need a specific pose, a specific silhouette, or your own sketch respected, you want a control layer: a depth map, a pose skeleton, an edge map pulled from a photo. Tools like ControlNet, img2img with a strength slider, and regional prompts are how professionals stop rolling dice. If any of that is new territory, the practical guide to creating art with generative tools walks through the workflow at a sensible pace.
Where it still breaks
Four things go wrong more than anything else. Hands, because fingers are high-frequency detail that the model has no structural reason to count. Text, because letterforms need exact shapes rather than plausible ones. Repeating patterns, because symmetry is hard to fake convincingly. And consistency, because the same character in six different poses will drift unless you build a reference system.
Science and history get slippery too. Ask for a specific species of bird in flight and you may get a beautiful creature that has never existed. This is the part of the process where a human eye still earns its keep, and it is covered well in this honest look at where AI drawing breaks and how artists work around it.
Where AI painting earns its place in a real workflow
Forget replacing finished paintings. The value sits earlier in the process.
A client wants a poster and cannot describe what they want. Generating forty rough directions in an hour gives you something to point at, and the conversation moves from abstract to specific in ten minutes. Background plates for animation, texture bases for 3D work, mood boards that used to take a full day of image searching. All of it compresses.
Mood in particular is where a well-written prompt shines. Ask for a rain-streaked window at 2 a.m., condensation on the glass, one lamp reflected in the wet street, and you get atmosphere that would take an afternoon to photograph. If that kind of scene-building interests you, the walkthrough on generating a 2 a.m. coffee shop that doesn’t exist is a good template to steal from.
What AI cannot do is tell you which of your forty options is the right one. That remains the job.
Money, credit, and the legal grey zone
Can you copyright an AI painting?
In the United States, mostly no. The Copyright Office has been consistent since 2023 that human authorship is required, and its 2025 report concluded that prompts alone do not give you enough control to claim the output. Purely machine-made images have been rejected in court. Registration has been granted where a human contributed substantial arrangement or editing, but the protection covers that contribution, not the whole picture.
Can you sell it?
Usually yes, with conditions. Adobe Stock accepts AI work with labelling. Getty largely does not. Etsy requires disclosure in most categories. Print-on-demand platforms vary wildly and change their policies often enough that checking before you upload is worth the two minutes.
The practical move is disclosure, early and plainly. Clients rarely mind. What they mind is finding out later.
A practice routine that actually builds skill
Thirty focused minutes a day beats a four-hour weekend binge, because the skill being built is judgment, and judgment needs repetition with feedback.
Spend five minutes choosing one real painting you admire and writing down what makes it work. Ten minutes translating that into a prompt and generating. Ten minutes picking the closest result and editing it by hand, in Photoshop or Procreate or on paper. Five minutes writing a single sentence about what went wrong.
Keep a folder of the failures. Six months from now it will be the most instructive thing you own, and the gap between what you could generate in January and what you can generate in July will be obvious in a way that daily practice never feels like at the time.