Search for “AI prompts” and most of what comes back is advice about prompts rather than prompts. You get told to be specific, to add context, to iterate. All true, all aimed at chatbots, and none of it puts a single line of text in your clipboard.
So this article is mostly prompts. Complete ones, with the swappable part in square brackets, grouped by what you are trying to make: portraits, landscapes, funny stuff, couples, Halloween, Christmas, character sheets, edits to photos you already have, and short video clips. Each one names a model that handles it well.
You can try any of them on the free AI image generator without signing up — guests draft on DaVinci2 and Z-Image Turbo, and the other models named here need an account. A standard image costs 4 energy, so a round of four drafts costs 16 — worth knowing before you burn a 50-energy model on a prompt you haven't tested yet.
Read the anatomy section first if you want to stop copying and start writing your own. Skip to the prompts if you came here to make something tonight.
- Anatomy
- Portraits 01–06
- Landscapes 07–11
- Funny 12–16
- Halloween & Christmas 17–22
- Couples & characters 23–28
- Photo edits 29–33
- Video 34–37
- Broken renders
- FAQ
What an image prompt is actually made of
Text-prompt guides describe five components: task, context, role, format and constraints. That works for a chat request. An image model needs different slots, because it isn't answering you — it's rendering a scene, and it will invent anything you leave unspecified.
Six slots do almost all the work:
a woman in her sixties with cropped grey hair1970s colour film photograph · ink and watercoloursingle work lamp overhead, deep shadows85mm lens at f/1.8, shallow depth of fieldfull body, centred, neutral backdropmuted earth tones · teal and amberThen aspect ratio, which is a setting rather than a word:
Order matters. Deep Dream Generator's own prompt-writing guide recommends arranging details in descending order of importance, most important first, and that matches what we see in practice: words near the front survive, words at the end get treated as suggestions.
Length has a working range. Under about 15 words and the model fills the gaps with its house style. Over about 70 and details start dropping out silently — you can't tell which ones until you compare renders.
Full sentences work better on newer instruction-following models like Nano Banana 2, ChatGPT 2 and SeeDream; comma-separated keyword stacks still work better on Stable Diffusion lineage models.
Portrait and people prompts
Portraits fail on lighting, not on subject description. Name one light source and where it comes from and the render steadies immediately.
- SeeDream · 30 energy
Close-up portrait of [a woman in her sixties with cropped grey hair], soft window light from the left, shallow depth of field, 85mm lens at f/1.8, muted earth tones, plain grey backdrop, 4:5
- Flux 2 · 25 energy
[A blacksmith] photographed in [his workshop], full body, 35mm lens, single work lamp overhead, deep shadows, sparks in the air, documentary photography, 3:2
- Nano Banana 2 · 50 energy
Studio headshot of [a young man in a black turtleneck], one softbox at 45 degrees, black seamless background, catchlights visible in both eyes, sharp focus on the near eye, 1:1
[A market trader] mid-sentence, hands raised, caught in movement, direct on-camera flash at night, hard shadow behind, grainy 400-speed film, 3:2
Painted portrait of [a violinist] in the manner of 17th-century Dutch oil painting, single candle from below, dark umber background, visible brushwork and craquelure, 4:5
Overhead shot of [a chef's hands] folding dough on a floured steel counter, cool daylight, no face visible, top-down, 1:1
Swap what's in the brackets and leave the rest alone. That is the whole point of a prompt library — the lighting and camera language is the part you paid for, and it survives a change of subject.
Landscape, place and wallpaper prompts
[A basalt coastline] under [a breaking storm], long exposure, water smoothed to silver, single lighthouse on the right third, cold blue-grey palette, 3:2
[A pine forest] in fog at first light, layered depth, telephoto compression at 200mm, desaturated greens, no sky visible, 4:5
Empty [1950s American diner] interior at 3am, fluorescent tubes and one neon sign as the only light, wet reflections on the linoleum, 2.39:1
Aerial view straight down on [terraced rice fields], geometric pattern, midday light, high contrast, minimal composition, 16:9
- Z-Image Turbo · 10 energy · cheap drafts
[A single tree] on a snow field, everything else empty white, tiny subject in a large frame, flat overcast light, 16:9
Funny prompts, and why “funny” is the wrong word to use
Asking a model for something funny reliably gets you clutter: clowns, exaggerated expressions, cartoon slapstick. The comedy comes from treating an absurd subject with a completely straight face, so describe it as if it were a serious commission.
A pigeon sitting behind a desk in a corner office, wearing a tiny headset, city skyline through the window, corporate stock photography lighting, 50mm, entirely deadpan, 3:2
Museum oil painting of [a knight fighting a photocopier], cracked varnish, gilt frame edge visible, raking gallery light, 4:5
Nature documentary still: [a cat] photographed from a hide with a 600mm lens, shallow focus, long grass in the foreground, as if a rare predator, 16:9
Formal Victorian family portrait, everyone seated and solemn, except [one large dog] in the centre chair, sepia albumen print, soft vignette, 4:5
Product launch photograph of [a single boiled potato] on a rotating plinth, black studio, rim lighting, smoke haze, 1:1
Halloween and Christmas prompts
Seasonal prompts drift towards greeting-card generic. The fix is to fix the light source, because both seasons are essentially lighting problems — a flame below the face, string lights behind it.
Halloween · a flame below the face
Selfie of [me] holding a carved pumpkin lantern on a suburban porch at dusk, lantern light from below as the only light source, orange and deep blue palette, 35mm phone camera look, slight motion blur, 4:5
Victorian ghost-story illustration: [a child in a nightgown] at the top of a dark staircase holding one candle, ink and watercolour, soot black and bone white only, 2:3
[An empty country road] at night, fog, one streetlight, long shadows stretching toward the camera, nothing else in frame, 2.39:1
Christmas · string lights behind it
Family portrait of [four people] in knitted jumpers beside a decorated tree, warm tungsten string lights as the key light, heavy film grain, 1970s colour photo, 3:2
Studio product shot of [a stack of wrapped presents] on pale grey seamless paper, soft top light, one red ribbon as the only saturated colour, 1:1
[A lit shop window] on a snowing street at night, seen from outside across the pavement, warm interior against cold blue exterior, reflections in the glass, 16:9
For selfie-style seasonal images, models that accept a reference photo matter more than the wording. SeeDream can take a face and likeness from an uploaded reference and place it into the described scene, which is what makes prompt 17 work as a picture of you rather than a stranger.
Couples, characters and the POV format
Two faces in one frame doubles the chance of an artifact. Our most reliable couple prompts either push both faces small in the frame or crop them out entirely.
[A couple in their thirties] walking a wet city street at night under one umbrella, seen from behind, backlit by shop signs, rain streaks visible in the light, teal and amber, 2.39:1
Two hands on a café table, one resting on the other's wrist, top-down, natural window light, 50mm, out-of-focus cups in the corner, 1:1
[A couple] silhouetted against [a sunset lake], tiny in the frame, backlit, warm orange sky, no facial detail, 16:9
Close dance-floor portrait of [two people] mid-laugh, direct flash, dark background, 1990s party photograph, motion blur at the edges, 4:5
For characters you intend to reuse, write a sheet rather than a scene, then keep the seed:
Character design sheet: [a desert courier in a solar-panel cloak], three-quarter view, full body, flat colour, thick linework, neutral background, 3:2
[The same character], front view and side view side by side, identical outfit and colours, plain white background, 16:9
The POV format is the one prompt shape that still surprises people. The base line:
-
Create an image of me [insert any adventure here], but make it from my POV. Remember to keep the image POV.
The repeat of “keep the image POV” is what holds the first-person perspective instead of drifting to a third-person shot of you; there are twelve worked variations and a selfie-reference version if you want to see the pattern stretched.
Editing prompts for a photo you already have
Editing prompts are instructions, not descriptions. The failure mode is the model rebuilding the whole picture when you asked it to change one thing, so every edit prompt should name what must stay identical — the underlined clause in each prompt below.
Replace the background with [a plain studio grey sweep]. Keep the subject's edges, hair detail and cast shadow intact. Match the original lighting direction.
Change [the red jacket] to [a beige canvas jacket]. Leave the pose, face, hands and background completely unchanged.
Restyle this photograph as [a 1950s hand-tinted postcard]. Keep the composition and all facial features unchanged. Do not add or remove objects.
Remove [the parked car] on the left and reconstruct the wall behind it, matching the existing brickwork and shadow.
Extend this image to [16:9] by continuing the scene left and right. Match grain, colour and perspective. Add no new subjects.
On the DDG side the AI Image Editor runs 15–50 energy depending on the model you pick, Remove Background is 15, and the AI Image Upscaler at 2K is 10. Upscale last, always — upscaling a draft you then re-edit is energy spent twice.
Prompting for video: one motion per clip
Video prompts break when they contain two moving things. Name the camera move, name the subject move, then say what stays still. Four to six seconds holds roughly one action.
Image-to-video · start from a still you already like
The camera pushes in slowly on [the subject]. [Her hair moves in a light breeze.] Background stays static. No cuts.
Slow orbit to the right around [the object]. Lighting stays constant. Nothing else in the frame moves.
Text-to-video
Handheld shot following [a cyclist] through [a wet night market], camera trailing at shoulder height, neon reflections on the road, single continuous take, no cuts.
Static locked-off wide shot of [a kitchen table] as [steam rises from a cup]. Nothing else moves. Natural window light.
Image-to-video gives you more control for less waste, because you have already approved the first frame. Text-to-video gives you movement the still would never have implied. Cost separates the models sharply:
Energy for one short clip
Sound off. Turning audio on doubles Seedance and Kling, and adds half again to Veo. A draft still is shown for scale.
Draft the still cheaply, animate once — in the video generator.
When the render comes back broken
Every one of these has a wording fix, not a settings fix.
| Problem | What to change |
|---|---|
| Mangled hands | Crop them out, or specify the position: “hands in pockets”, “one hand resting flat on the table” |
| Garbled lettering | Put the exact text in quotation marks and use a model built for type — Ideogram v4 (40 energy) is the one we go to |
| Wrong number of subjects | Move the count to the first four words: “Three people, seated…” |
| Perspective drifting | Repeat the perspective at the end of the prompt, the way the POV prompt repeats “keep the image POV” |
| Exclusions ignored | Rewrite them as positives. “No clutter” fails; “bare wooden table, empty background” works |
| Style overwhelming the subject | Cut the style words to two and see what the subject looks like alone |
Negative prompts help most on Stable Diffusion lineage models, where a short list — watermark, text, extra limbs — genuinely suppresses those features. On instruction-following models, a negative list often reads as another mention of the thing. Describe the frame you want instead.
The same prompt is also not the same prompt across models. A comma stack that produces a clean render on Flux Klein (8 energy) can come out literal and flat on ChatGPT 2 (30 energy), which prefers a sentence. When you move a prompt to a new model, keep the subject and lighting, then rewrite the grammar around them.
Where to find prompts that already worked
The fastest way to learn prompt phrasing is to read prompts attached to results, not prompts in a list. On Deep Dream Generator, each image's page shows the prompt that made it — with its modifiers when you're signed in — and a Try button loads it into the generator so you can change one thing. The #prompts tag is a reasonable place to start; neighbouring tags are large, with #landscape at 97,572 items and #art at 72,119.
Working backwards from an image also works. The AI Art Assistant turns an uploaded image into a detailed text-to-image prompt covering subject, style, lighting and mood, polishes a prompt you already have, and can be tried without an account. The standalone Image to Text tool costs 10 energy per run.
Frequently asked questions
What makes a prompt “good” rather than just long?
A good prompt is stable — it produces a strong result reliably, across several seeds, without you rerolling ten times. Stability comes from filling the six slots once each, in order of importance, and resisting the urge to add a tenth adjective. If two words in a prompt fight each other, the model picks one and you never find out which.
How long should an image prompt be?
Twenty-five to sixty words covers almost everything in this article. Below that you're delegating decisions to the model; above roughly seventy words, details begin dropping out without warning. If you need more specificity than that, put it in an edit prompt on the second pass rather than in the first prompt.
Do I need to pay to run these prompts?
No. Deep Dream Generator lets you start generating for free without signing up, and free users receive energy for images. Paid plans start at $9 a month for the Basic tier with 750 monthly energy, and go up through Advanced at $19, Professional at $39 and Ultra at $99; the plans page lists the energy cost of every model so you can price an experiment before you run it.
Which model should I pick for the money?
Draft on something cheap — DaVinci2 at 4 energy, Flux Klein at 8, Z-Image Turbo at 10 — until the wording is settled. Then spend on the final render: SeeDream at 30, Flux 2 or Grok at 25, Ideogram v4 at 40 for anything with readable type, Nano Banana 2 at 50. Deciding composition on a 50-energy model is the most common way people run out of energy early.
Do I own the images a prompt produces?
You retain rights to the images you create, and downloads come without watermarks at full quality. Commercial use is covered for images made while you had a paid monthly subscription or with energy from an energy pack, so check that against your intended use before you publish a client job.
Should I name a living artist in a prompt?
We'd advise against it, and not only for legal reasons: it's imprecise. “In the style of [living artist]” hands the whole decision to the model's idea of that person. Describing the actual technique — palette, brushwork, lighting, medium, era — gets you closer and keeps the result yours. The same caution applies to real people's likenesses.
Can I put my own face in a scene?
Yes, with a reference image. SeeDream can take your face and likeness from an uploaded photo and place it into the described scene, which is what the selfie variant of the POV prompt relies on. Keep the reference well lit and roughly front-facing; a dim or heavily angled photo pushes the model back toward a generic face.
Why does the same prompt look different every time?
The seed changes. If you get a composition you want to keep, note the seed and reuse it while you change one word at a time — that's the only way to tell whether your edit improved the prompt or you simply drew a luckier number.
A prompt library is scaffolding
The prompts above are real ones with the variables marked so you can lift them tonight, but the part worth keeping is the lighting and camera language inside them — that travels to subjects nobody has written a prompt for yet.
Start with one from the list, change a single slot, and run it twice at 4 energy before you spend 50. Then go and read the prompt under somebody else's render in the gallery. That loop — copy, change one thing, compare — is the entire skill.
Comments (5)