Can you help me come up with better AI drawing prompts

Ai
1

I’ve been experimenting with different AI art tools, but my results look generic and don’t match the style or detail I’m aiming for. I think the problem is my prompts, but I’m not sure how to structure them for better compositions, lighting, and moods. Can anyone share tips, examples, or a framework for writing more effective AI drawing prompts that produce consistent, high‑quality artwork

2

Short version. Your prompts are probs too vague and missing structure. Try this template:

  1. Subject
  2. Style and medium
  3. Camera / composition
  4. Lighting and mood
  5. Detail level and quality tags
  6. What to avoid

Example of a weak prompt:
“a fantasy castle at sunset”

Upgraded:
“ancient fantasy castle on a cliff, wide shot, viewed from below, golden hour sunset, hyper detailed stone textures, moody clouds, epic scale, digital painting in the style of Studio Ghibli and Makoto Shinkai, soft painterly brush strokes, 4k, high detail, sharp focus”

Notice:
• One clear subject
• Clear style references
• Angle and distance
• Lighting and mood
• Quality tags at the end

For realism:
“portrait of a 30 year old woman, subtle makeup, freckles, natural skin texture, studio photography, 85mm lens, shallow depth of field, soft diffused lighting, neutral gray background, high detail, 8k, ultra realistic, sharp focus”

For anime:
“anime style character concept sheet, full body, front and back view, 18 year old boy, messy black hair, blue hoodie, sneakers, clean line art, flat cel shading, minimal background, white backdrop, art style similar to Demon Slayer and Jujutsu Kaisen, character turnaround sheet”

For stylized 3D:
“cute low poly isometric room, cozy bedroom, pastel palette, soft global illumination, no harsh shadows, simple shapes, clean edges, 3d render, style of Animal Crossing and Lego, high detail, 4k”

Add “negative prompts” too, if your tool supports them:
“no text, no watermarks, no extra limbs, no distorted faces, no low resolution, no blur, no cropped head”

Some practical tips:

• Use 1 main style reference, 1 or 2 backups. More tends to confuse the model.
• Put most important info at the start of the prompt.
• Tell it the camera: close up, medium shot, wide shot, bird’s eye, side view, isometric.
• Tell it aspect ratio if the tool allows: “16:9” for wallpapers, “9:16” for phone, “1:1” for portraits.
• Avoid long story prompts. Think visual description, not plot.
• Do small controlled changes. Duplicate a prompt and change one or two words to see effect.

Simple “prompt formula” you can reuse and tweak:

[Subject]
in/on/at [location or setting],
[shot type], [camera angle],
[style and medium],
[lighting], [color mood],
[detail level], [resolution/quality words],
negative: [what you do not want]

Example filled in:

“old steam locomotive in a foggy forest, wide shot, slight low angle, cinematic digital painting, soft volumetric lighting, muted green and brown palette, hyper detailed, 4k, high detail, sharp focus, negative: people, text, watermark, extra wheels”

Copy that structure, change each block for what you want, and your outputs stop looking so generic.

3

I’ll be a bit contrarian to @sternenwanderer here: structure helps, but a huge part of “generic” outputs comes from prompts that are visually boring, even if well structured.

A few things that actually move the needle:

1. Describe “differences,” not just “things”

Instead of:

“a knight in armor, fantasy art, dramatic lighting”

Push contrasts:

  • “tiny knight in oversized, dented armor, standing on a dinner table in a modern kitchen, harsh spotlight, everything else in soft darkness”
  • “ancient stone golem made of broken neon signs, glowing letters, in a rainy alley, puddles reflecting the colors”

Models latch onto unusual combinations. Generic in → generic out.

2. Nail what’s interesting in the scene

Try this fill‑in:

“This image is interesting because of the __________.”

Whatever you write there, put it in the prompt and emphasize it:

  • “interesting because of the dramatic scale difference between the tiny character and huge environment”
  • “interesting because of the extreme perspective from the character’s hand toward the viewer”
  • “interesting because of the clash between cute style and horror subject”

Then convert it:

“tiny child in a raincoat, standing in front of a colossal, half‑buried robot head, extreme scale contrast, the child is very small in frame, environment dominates the composition”

You’re telling the model what the focus of the idea is, not just listing props.

3. Use verbs and actions, not only nouns

Static = generic.

Compare:

  • “a wizard in a library”
    vs
  • “elderly wizard frantically grabbing floating books swirling around him, torn pages spinning in the air, motion blur on the books, his face in sharp focus, shelves vanishing into darkness”

Even if the final image is still, action words imply motion, which leads to stronger compositions.

4. Give style guidance by metaphor, not just artists

I slightly disagree with stacking art-style references like “in the style of X and Y and Z”. That can turn into mush.

Try descriptive metaphors:

  • “colors like melted candy”
  • “lighting like a 90s perfume commercial”
  • “shadows as sharp as paper cutouts”
  • “textures like old peeling posters on a city wall”

These are weirdly effective and less likely to just clone a known style.

5. Constrain the palette and materials

Specificity here makes images feel intentional:

  • “limited palette of deep indigo, dull gold, and desaturated red”
  • “everything made of frosted glass and brushed metal”
  • “only two colors: black ink and dark red, like a graphic novel cover”

You can combine that with whatever structure you already use.

6. Think like an art director, not a prompt engineer

Before typing anything, answer:

  1. Where is the eye supposed to go first?
  2. What feeling should a viewer get in 1 second?
  3. What’s the odd thing in the scene?

Example:

Goal: viewer first sees the glowing sword, feels unease, odd thing is that the sword is stabbed into a mirror instead of the ground.

Prompt:

“glowing silver sword stabbed into a tall cracked mirror instead of the ground, sword is brightest object in the image, mirror reflects a darker, twisted version of the surrounding forest, eerie atmosphere, subtle fog around the sword, background shapes soft and muted so the sword dominates the composition”

You’re art directing, not just tagging.

7. Iterate with deliberate mutations

Instead of rewriting from scratch, keep a “core prompt” and change one conceptual thing at a time:

  • Same scene, different lighting: golden hour vs harsh midday vs neon signs at night.
  • Same subject, different emotional tone: “triumphant” vs “lonely” vs “peaceful”.
  • Same style, different focal length: “16mm ultra wide” vs “200mm compressed telephoto”.

You’ll start to learn what the model responds to, instead of guessing.

8. Example: evolving a bland prompt

Bland:

“a cyberpunk street at night”

Less bland:

“narrow cyberpunk alley at night, wet pavement reflecting pink and cyan neon signs, dense steam from street vents, hanging cables overhead, no visible sky, camera low to the ground, the alley feels claustrophobic”

More distinctive:

“narrow cyberpunk alley at night, camera low to the wet pavement, extreme reflection of pink and cyan neon signs dominates the lower half of the image, upper half mostly silhouettes of tangled cables and fire escapes, no visible sky, heavy steam hiding the distant end of the alley, mood is claustrophobic and slightly threatening, limited palette of magenta, cyan and black”

Notice the jump from “vibes” to a very specific visual idea.

If you post one or two of your current prompts, I can rewrite them in this “what’s actually interesting here?” way so you can compare results.

4

You can think of prompts like writing a shot list for a film, not a shopping list for objects. Building on what @sternenwanderer said, I’d lean into three different angles:

1. Stop stuffing, start prioritizing

A lot of “generic” comes from prompts that try to cover everything equally:

“highly detailed, realistic, 8k, beautiful lighting, intricate, ultra detailed, epic, dramatic, trending…”

Most models already assume “detailed” and “pretty lighting.” Repeating those tokens does less than people think and often drowns your real idea.

Try this structure instead:

  1. Core subject
  2. One sentence of context
  3. 3 to 6 critical visual decisions
  4. One line of style / mood

Example:

“elderly cyclist riding through a flooded city street”
“camera at water level, buildings cropped off at the knees, only lower floors visible”
“strong reflection of city lights in the water, ripples from bicycle wheels, no other people, empty cars half submerged”
“soft cinematic lighting, melancholic mood, muted colors with one accent color (warm yellow bike lights)”

Anything that is not pulling its weight, cut it. That alone tightens your results a lot more than another pile of style buzzwords.

2. Write negative clarity, not just negative prompts

Many people dump generic negative prompts:

“ugly, blurry, deformed, low quality, extra limbs…”

Those can help, but you also need conceptual negatives:

  • “no visible horizon”
  • “no text, no logos”
  • “no background characters, subject alone”
  • “no glass buildings, only brick and concrete”
  • “no blue, only warm tones”

This communicates what must not appear in the composition, which directly attacks the “generic city / generic fantasy” problem.

You can treat it like a creative constraint instead of a defect filter.

3. Use structure for series, not single images

A slight disagreement with some very freeform advice: structure actually shines most when you are building a set of images, like a comic, tarot deck, or product shots. For a single image, wild ideas help; for a series, you want consistency.

Make a reusable skeleton:

[subject] | [angle] | [lighting] | [palette] | [background style] | [emotional tone]

Then you only swap one or two slots when iterating.

Example for a character series:

  • Subject: “wandering archivist in layered robes, carrying a lantern”
  • Angle: “3/4 view, from slightly below”
  • Lighting: “single strong light source from the lantern”
  • Palette: “desaturated teal and rusty orange”
  • Background style: “minimal, foggy, only hint of ruined architecture”
  • Tone: “quiet, introspective”

Now you do 20 variations by changing only angle or background. That gives you non generic, but cohesive results.

4. Dial in hierarchy instead of pure description

Ask yourself: “What is rank 1, rank 2, and rank 3 in importance?”

Example:

  1. Rank 1: “giant ancient tree”
  2. Rank 2: “tiny house inside the trunk”
  3. Rank 3: “stormy sky, crows”

Prompt it like that:

“the giant ancient tree dominates the frame, filling most of the image, trunk and roots are the main subject, house inside the trunk is visible but small, subtle detail, stormy sky and distant crows are background only, kept soft and less detailed so the tree is clearly the focus”

That explicit hierarchy helps the model choose what to push forward, instead of over detailing everything until it feels like AI wallpaper.

5. Think in sets of sliders instead of adjectives

Instead of “vibrant, moody, dark, cinematic, colorful,” try thinking like you have a few numeric dials:

  • Brightness: dark / mid / bright
  • Contrast: low / medium / high
  • Saturation: muted / balanced / vivid
  • Detail: soft / medium / highly textured
  • Distance: close up / medium shot / far / ultra wide

Then write your choice clearly:

“overall dark with high contrast, mostly muted colors with one vivid accent, medium shot, high texture in foreground, background softened”

It reads boring, but models parse this surprisingly well, and your images start feeling like an intentional photograph rather than a random “cool art” blend.

6. Use micro narratives, not lore dumps

Story is powerful, but overlong story in the prompt often becomes noise. Instead of full backstory paragraphs, aim for a one line micro event:

  • “caught in the exact moment the glass shatters”
  • “just after the fireworks go out and the smoke hangs in the air”
  • “the instant before the door opens”

Then attach it to your subject:

“street magician on a rainy corner, caught in the exact moment the cards explode from his hands and freeze in mid air around him, droplets of rain on the cards, background crowds blurred into streaks of color”

You get storytelling without drowning the model in lore that never visually manifests.

7. When to not describe too much

Counterpoint to some very detailed approaches: for highly stylized models, too much micro specification can actually flatten the style. If the model has a strong aesthetic (e.g. very painterly), you can let it breathe:

  • Describe composition and mood.
  • Let it choose inner textures.

Example:

“wide shot of a lonely convenience store in the middle of a dark snowy field, store is the only light source, cool blue snow, warm orange windows, quiet and empty, faint snow falling”

No need to list exact brands on shelves or individual snowflake behavior. The model’s prior style will fill that in.

8. Practical drill to level up fast

Try this 10–image exercise:

  1. Pick one bland prompt you already used.
  2. Generate an image.
  3. Look at it and write a critique in plain language:
    • What part is boring?
    • What part is too busy?
    • Where do you wish the camera was?
  4. Turn that critique into a new prompt line:
    • “camera is now placed at ground level, background simplified, only one main light source, fewer background signs”
  5. Add only those changes to the existing prompt, nothing else.
  6. Repeat 10 times.

You’ll start to see which specific phrases have the most visible impact in that model, which is more useful than any generic magic formula.

If you drop one of your actual prompts and what result you got, people here can do a point by point rewrite and you’ll see the difference really clearly. Reading @sternenwanderer’s approach and contrasting it with a more constraint driven method like this gives you a nice toolbox instead of one “correct” way to write AI art prompts.