← All guides

How to write AI image prompts that Nanomelon can follow

Craft notes for the homepage prompt box · text-to-image and image-to-image

A prompt is not a poem and it is not a search query. It is a brief for a system that will fill every unspecified gap with a statistically common guess. If you write “a girl, with blonde hair, radiating sunshine,” the model must invent age, pose, camera distance, clothing, background, and whether “sunshine” means a weather report or a lighting setup. Sometimes that guess is charming. Usually it is generic. This page is how we brief the same models from the Nanomelon homepage so the guessable parts shrink.

The five-part brief

Write the prompt in this order, even if you later shuffle the sentences. Order is not magic; completeness is.

  1. Subject. Who or what is in focus, with one or two visual facts that matter (age range, species, garment, product type). Avoid celebrity names and brand mascots; those triggers often fail the policy filter and waste a click even when credits are refunded.
  2. Action or pose. Standing, seated, mid-stride, looking at the camera, hands occupied. If you skip this, you get a catalog stare.
  3. Setting. Interior or exterior, time of day, a few props that belong there. “Studio” is a setting. “Aesthetic background” is not.
  4. Light and camera. Direction (backlit, window left, overcast), quality (soft, hard, neon), and a lens hint if you care about distortion or compression (35mm environmental, 85mm portrait). You do not need to fake EXIF. You need to stop the model defaulting to beauty-dish glamour on every face.
  5. Medium and finish. Photograph, watercolor, woodblock, 3D render, comic ink. If you want a painting, say so; otherwise many models will give you a sharp digital photo because that dominates the training mix.

Example of a short prompt that still has all five parts: “A child in a red festival jacket rides a horse-shaped paper lantern, night street in an old Chinese town, warm lantern light from inside the paper, watercolor illustration, 3:4.” Compare that with the public example on gallery item 105, which adds fireworks, sky lanterns, and architecture so the model has more to lock onto.

What to leave out

Weight words like “masterpiece, 8k, trending on artstation, ultra detailed” rarely buy you structure. They compete with the nouns. If the hands are wrong, adding “highly detailed hands” sometimes helps and sometimes gives you six fingers with more wrinkles. Prefer a clearer pose: “hands in pockets” or “holding a paper lantern with both hands.”

Negative prompting, if your selected model exposes it, is for classes of errors you have already seen: extra limbs, watermark, text overlays, a second head. It is not a place to dump the opposite of your story. “No ugly, no bad anatomy” is superstition. “No watermark, no caption text” is a real constraint because models like to sign images.

Do not paste another site’s entire style block unless you understand each clause. A 120-word prompt that repeats “cinematic lighting” three times is thinner than a 40-word prompt with one lighting sentence.

Aspect ratio is part of the prompt

On Nanomelon you choose ratio in the form, not only in the text. 1:1 suits icons and avatars. 3:4 and 4:5 suit posters and character sheets. 16:9 suits environments and banners. If you ask for a “full body in landscape 16:9,” the model will invent a lot of empty floor. If you want a face, do not pick 16:9 and hope. Matching the ratio to the crop you actually need is the cheapest quality upgrade we have.

Iterate like an editor, not a slot machine

Change one axis per rerun: light, or wardrobe, or camera distance. If you rewrite the whole paragraph every time, you cannot tell which clause caused the better picture. Use the same model while you learn its habits; switching engines mid-test mixes two sets of defaults. When a result is close, switch to image-to-image and describe only the delta (“keep the pose, replace daylight with lantern-lit night”).

Public gallery cards that used text2img are the right place to steal structure. Copy the skeleton (subject → place → light → medium), then replace the nouns. Copying the entire string and changing one adjective is how you get a site full of near-duplicates; it also teaches you nothing.

Style words that actually steer

Medium names (gouache, risograph, charcoal on newsprint) are stronger than mood adjectives (epic, stunning, beautiful). Cultural references to a print tradition (“Qing-era new-year print, flat color, decorative border”) are stronger than “Chinese style,” which can mean anything from ink wash to neon Shanghai. If you need a specific festival, name the objects: horse lantern, sky lantern, couplets, not “traditional vibe.”

For portraits, skin should be described as a lighting problem, not a beauty filter: “soft window light, visible pores, no beauty-blur.” For products, describe the surface: brushed aluminum, condensation on glass, cardboard edge wear. Models over-polish unless you ask for wear.

Policy-safe prompting

Nanomelon and the upstream models will refuse sexual content involving minors, some public-figure likenesses, and some trademarked characters. If you hit a policy message, credits are not deducted. Rephrasing the same disallowed request is a waste of your time and may flag the account. Write original characters, or use image-to-image on photos you have the right to use.

When you publish a work to the public gallery, assume a stranger will read the prompt. Do not put emails, real addresses, or other people’s private photos in the text.

A checklist before you click Generate

If four of those are yes, generate twice at quantity 1 rather than five at once. Compare, then spend the next credits on the winner’s cousin. That is how a small free allotment still teaches the craft. For cost and model choice, continue with models and credits.

Open the Nanomelon generator and paste a five-part brief. Study a finished example in the festival lantern gallery page or the backlit portrait page.