Prompting for game sprites: what to say and what to leave out
Sprite prompting is its own small skill, and it's mostly about restraint. The model already knows what pixel art is. Your job is to pin down the four things it can't guess, and stop talking before you confuse it.
The structure that works
Four slots, in roughly this order:
[subject with 2-3 concrete details] + [pose or action] + [view] + [style anchor]
Examples:
armoured knight with a tower shield, standing idle, facing right, 16-bit RPG style
red health potion in a round glass bottle, cork stopper, front view, clean modern indie pixel art
Structured prompts, real outputs. The potion is the example above, verbatim.
- Subject details beat adjectives. "Tower shield, horned helmet, torn red cape" gives the model shapes to draw. "Epic, detailed, awesome" gives it nothing.
- One pose, named. Without a pose you get the model's default, which is usually a stiff A-pose. "Mid-swing", "crouching", "hovering" instantly improve a sprite.
- Say the view. "Side view", "front view", "top-down". Game sprites have a camera convention; the model needs to know yours.
- Anchor style with names, not vibes. Era references ("16-bit", "SNES-era") and game names work far better than adjectives. We've found "isometric" alone barely works, but "isometric, like Stardew Valley furniture" snaps the model into the right projection. Concrete references are the single biggest lever in sprite prompting.
What to leave out
- Resolution talk. "32x32", "low res", "tiny sprite" makes models draw chunkier fake pixels, not smaller sprites. Generate at full detail and downscale properly afterwards (see why). Detail level and output size are separate decisions.
- Background descriptions. You want a transparent-background sprite. Describing a scene ("in a dark forest") invites the model to paint one, and now you're cropping your character out of an environment.
- Multiple subjects, accidentally. "A knight fighting a dragon" is a scene, not a sprite. If you want both, generate them separately, or use a sprite pack mode to get a matched set deliberately.
- Words you want rendered as text. Models love writing labels on things. If your item doesn't need lettering, don't mention any.
The failure modes to know
- The background eats the subject. Tools that key out a solid background colour have a classic bug: ask for a blue dragon, get a blue background, and the keying removes half the dragon. Good tools guard against this behind the scenes (SpriteLab does). If you're prompting a raw model yourself, ask for a solid background in a colour that appears nowhere on your subject.
- Style drift across a set. Ten separate generations gives you ten interpretations. For matched sets, generate together (one grid call) rather than separately; see the sprite sheets guide.
- Over-prompting. Past about 30 words of subject description, additions start fighting each other. If a long prompt is failing, the fix is usually deletion.
Iterate cheaply
The biggest waste in AI sprite work is re-rolling for things that aren't the model's job. The look of a sprite (colours, contrast, outline weight, palette) is post-processing, and in a tool with a proper tuning panel those changes are free and instant. Re-generate when the drawing is wrong: pose, proportions, missing details. Retune when the look is wrong. Knowing which is which roughly halves what you spend.
Test a prompt right now
Try the structure above and judge the first generation. Takes about 20 seconds.
TRY SPRITELAB FREE 25 free credits + 5 daily. No card needed.