"lone lighthouse keeper on a basalt cliff at dusk, storm rolling in from the left, warm lamp glow against cold teal sea, painterly gouache texture, generous empty sky for title, vertical 2:3 --ar 2:3 --style raw" A prompt like that, written for a hypothetical fantasy novel cover, is a small part of the job. Everything around it took longer: working out what the author wanted, binning dozens of pretty but wrong images, rebuilding the keeper's hands, and handing over a file the printer would accept. That whole loop is the work in these online ai image generation artist jobs, done part-time from a home desk anywhere in the world, at $22 per hour (USD).
Clients bring ads, book covers, social posts and the odd album sleeve. You write and refine prompts in Midjourney and Stable Diffusion until a few strong options exist, help the client pick one, then repair and finish it in Photoshop. Upscaling and export close it out. That is the real ai image generation artist job description: the person in this seat is closer to a production artist than a button presser, and repeat clients are the ones who got a clean file without asking twice.
Most requests arrive half formed. "Something moody, like that Netflix poster" is common. Before generating anything, pin down final size and aspect ratio, where text will sit, print or screen, brand colours, and anything the client hates. A cover needs quiet space at the top for a title and at the spine edge for trim. An Instagram carousel needs every frame to read at thumbnail size. Sending that checklist back as questions usually saves a whole revision round.
Midjourney is quick for exploring mood and composition. Run a few loose prompts, look at the grid, and change one variable at a time: light direction, lens feel, palette, medium. Style references and character references help keep a series consistent. Stable Diffusion is where control lives. With a checkpoint chosen for the look, a LoRA for a specific style or product, and ControlNet for pose or depth, you can hold a composition steady while you refine the details. Seeds matter. Write them down, because a client will ask for "the third one, but with the coat in red" a week later.
Good prompt skills come from looking, not memorised magic words. Know what "rim light", "three-quarter view", "overcast soft shadow" and "risograph grain" actually look like, and you will describe images precisely.
Send three to five options, not thirty. Too many and the client freezes. Lay them on one board with short notes: "A: strongest silhouette, B: warmer, more romantic, C: closest to your reference." Ask which one, and what they would change. Then commit. Iterations after a direction is chosen should be refinements, not a fresh search, and your hourly estimate should say so up front.
This is the part people underestimate. Generators still produce six-fingered grips, melted earrings, buttons that drift across a seam, windows that do not line up, and lettering that looks like a language from a dream. Fixing them is ordinary retouching: generative fill or inpainting for a first pass, then clone, heal, liquify and some painting by hand to make it believable. Never trust generated text. Remove it and set real type, or leave clear space for the client's designer. Zoom to 100 percent and check every edge before you call it done; a printer will show what your laptop screen hid.
Raw generations are rarely large enough for print. Use a dedicated upscaler (Topaz, an ESRGAN model in ComfyUI, or Midjourney's own upscale) and then inspect the result, because upscalers invent texture and can turn skin waxy. For print, work out pixel dimensions from the trim size plus bleed at the resolution the printer asks for. For web, size to the platform and compress sensibly. Deliver what was requested, usually a layered PSD, a flattened TIFF or high-quality JPEG, and PNG or WebP for screens, with colour profiles noted and files named so a stranger could sort them.
Midjourney runs in the browser, so any decent laptop is enough there. Local Stable Diffusion through ComfyUI is different: it leans on an NVIDIA graphics card, and video memory decides what you can run. Older SD 1.5 models are forgiving on modest cards, while SDXL and newer models want considerably more VRAM, plus plenty of disk space for checkpoints. If your machine cannot cope, cloud GPU services let you rent time by the hour and run ComfyUI remotely. That works well for occasional heavy jobs, but count the rental against your rate. Tool subscriptions and hardware are normally your own cost as a freelancer, so price them in.
No degree is required here and no minimum years, so a thin CV will not shut you out. A portfolio does the talking. Show before and after pairs: the raw generation next to your retouched final, so a buyer can see the hands you rebuilt and the type you replaced. Group work by use (covers, product ads, social sets) rather than by tool. An eye for composition, lighting and colour matters more than any software trick, and that is learnable through drawing, photography or studying film stills.
AI art courses can speed things up, especially for ComfyUI workflows and ControlNet, and some offer a certificate of completion. Treat certificates as a bonus; clients will open your images before they read anything else. Free documentation, shared workflow files and daily practice on mock projects teach most of what a paid course does.
Image work like this gets posted and bought on marketplaces such as Fiverr, Upwork and Contra. Plenty of freelance ai image generation artist jobs start there, and ai image generation artist jobs for freshers do exist when the samples are good, even with no client history. Make narrow gig listings ("fantasy book covers with retouched hands and print-ready files") rather than "AI art of anything". Do two or three mock projects for invented clients and present them honestly as samples. Each marketplace sets its own policy on AI-generated work and disclosure, and policies shift, so reread them before publishing a gig. In part time ai image generation artist jobs like this one, the diary swings between packed weeks and empty ones, and no set hours come with it. The upside of ai image generation artist work from home is that a quiet week is a good week to build new samples.
On rate: this role pays $22 per hour (USD), and that is the ai image generation artist salary per hour you should plan around, not a promise of weekly totals. Track your time per stage on the first few jobs. You will learn quickly that retouching and revisions, not prompting, eat the clock.
Whether a client can use an image commercially comes down to what Midjourney, the model licence or the platform says this month, and which subscription tier you are on. Some tiers and model licences restrict business use, so check before every project and note what you used. Copyright protection for images made purely by a generator is limited or uncertain in many countries, and your own retouching or painting may change that picture in ways only a lawyer can assess. None of this is legal advice. The sensible habit is to tell clients plainly, in writing, which tools made the image and that exclusive ownership may not be guaranteed.
A digital illustrator builds an image from sketch to paint, mostly by hand. An AI image artist directs generation and then finishes by hand, so the strongest people in this work borrow the illustrator's habits: thumbnails, value checks, and the patience to repaint a hand rather than reroll it twelve times.