If you've ever tried to get a decent image for a game jam at 2am, you know the problem. You need forty assets, you have six hours, and none of your friends draw. Most tutorials assume you have weeks and a budget. This one assumes you have neither. Below is the workflow I've landed on for producing a usable 2D asset set with a text prompt and a browser, without the result looking like every other AI-generated game on itch.io.
The same approach works for tabletop tokens, visual novel backgrounds, and mobile puzzle sprites. I'll use a top-down forest tileset as the running example, but swap the nouns and it transfers.
Start with a style sheet, not a prompt
The mistake almost everyone makes is opening the prompt box first. Don't. Before you type anything, write down four decisions on paper: outline weight, palette size, light direction, and camera angle. Picking these up front is what keeps a set of forty images from looking like it came from forty different games.
For my forest set: two-pixel outline, twelve colors, light from the upper left, straight top-down. Those four lines become the tail end of every prompt you write for the next two hours. They matter more than the subject lines do, because subject variation is expected in a tileset and style drift is not.
Generate one hero image, then lock it
Pick the single most important asset in the set, the one players look at most. For a forest tileset that's a generic grass tile with a couple of trees. Generate it, iterate on it ten or fifteen times, and don't move on until you'd be happy shipping just that one.
Then save the prompt that produced it. Copy the exact wording, including the awkward bits. That saved prompt becomes the base for every other asset, and you only edit the subject clause when you move to the next tile. This is the single biggest lever on visual consistency, and it costs nothing.
Batch by category, not by asset
Generate all your ground tiles together, then all your props, then all your characters. Working in groups means you catch drift while it's still cheap to fix. If tile number thirty comes back with a chunky outline, you find out on tile number thirty of the ground batch rather than after you've already done the characters.
Lay the outputs out in a grid as you go. Squint at it. Anything that jumps out at you in a squint test is a style mismatch, even if you can't name it yet.
Fix the three problems AI actually has with game art
Symmetry. Default outputs are too centered and too symmetric to read as hand-made pixel art. Add "off-center composition, irregular edge" to prop prompts and it settles down.
Texture noise. AI loves busy detail that turns to mush when you scale down to actual sprite size. Generate at a larger size than you'll ship, then downscale and check whether the shape still reads. If it doesn't, the detail was never doing any work.
Palette creep. Each image introduces three or four new colors. After every batch, sample the images and pull them toward your twelve-color palette. A few minutes in an editor prevents the set from turning muddy.
If you're generating background plates or cutscene art rather than tiny sprites, the resolution question flips and it pays to know the model options available. aiimagegenerators.net is a reasonable place to compare them if you want to pick per-task instead of defaulting to one.
What to generate and what to hand-make
Here's the part nobody says out loud. AI is genuinely good at organic, one-off props: rocks, bushes, mushrooms, scattered debris, background clutter. It is bad at anything that needs to tile seamlessly, anything that needs to line up on a pixel grid, and UI elements.
So split the work. Generate your props and decorative layers. Hand-make your tileable ground, your character sprite sheets, and your UI. The seam problem alone justifies it, since a visible grid line across the whole level is the one flaw players notice immediately and never stop noticing.
Polish pass that makes it look intentional
Drop every asset into the same canvas size, align them to a consistent baseline, and add a single uniting element. For my forest set, that was a shared drop shadow with the same offset and opacity on every prop. One consistent shadow is worth more than an hour of extra generation, because it tells the eye that a person made these decisions.
Then do a light pass in any pixel editor: sharpen edges, clamp the palette, clean stray single pixels. Twenty minutes here separates "AI output" from "asset pack."
Keeping the set extensible
Before you call it done, save three things: the locked hero prompt, the style sheet lines, and a two-sentence note on your palette. Six months from now you'll want to add a rain variant or a seasonal recolor, and that note is the difference between a coherent addition and a set that suddenly looks patched. Version it in the same repo as your game files, not in a folder on your desktop.
That's the whole loop: decide style before prompting, lock a hero asset, batch by category, fix symmetry and noise and palette, hand-make anything that tiles, and unify with one shared treatment. You can pull it off in a weekend, and the result won't read as a generated set. When you need a quick second opinion on model choice or prompt phrasing, aiimagegenerators.net is a decent reference to keep open in a tab.