I’ve been experimenting with different AI image generators, but my results look generic or off-style compared to the stunning images I see others create. I’m trying to understand how people structure their prompts—things like style, lighting, camera angles, and keywords—to get sharper, more realistic, and consistent outputs. Could you share your best AI image prompts, plus any tips on wording that really improves image quality and creativity?
Short version. Your prompts are probably too vague, mixing styles, or fighting the model’s defaults. Here is what tends to work for realistic stuff.
I’ll use this structure:
SUBJECT + LOOK + CAMERA + LIGHT + COLOR + QUALITY + NEGATIVES
Example base prompt for realism:
“photo of a 28 year old woman, natural look, no makeup, slight smile, studio portrait, shot on 85mm lens, f1.8, shallow depth of field, soft diffused lighting, neutral color grading, ultra detailed skin, pores visible, 8k, high dynamic range”
Then add this kind of negative prompt:
“blurry, noisy, extra fingers, extra limbs, deformed hands, distorted face, asymmetrical eyes, crossed eyes, cartoon, illustration, 3d render, painting, low resolution, jpeg artifacts, watermark, text, logo”
Key tips:
-
Use photography language
Stuff like:
• “shot on 35mm / 50mm / 85mm lens”
• “f1.4, f1.8, f2.8, shallow depth of field, bokeh”
• “ISO 100, studio lighting, key light from the left, rim light”
• “close-up portrait, medium shot, full body shot”
This pushes models toward realism, not art styles. -
Control lighting
Prompt lighting clearly:
• “soft daylight, overcast, no harsh shadows”
• “golden hour sunlight, warm tone”
• “studio softbox lighting, even, no specular highlights”
Bad lighting words:
• Avoid “dramatic lighting” at first, it often gives crunchy or fake results.
• Skip “cinematic” if you want clean photo realism, it tends to make it stylized. -
Control style conflicts
Do not mix stuff like:
“photorealistic portrait, oil painting, anime style, 3d render”
Pick one. For realism:
Use words like “photo, photograph, realistic, DSLR, RAW photo, unedited look”. -
Use age, ethnicity, and setting
More detail, less guesswork:
Instead of “a person in a room”
Try:
“photo of a 35 year old Japanese man, casual outfit, sitting at a wooden desk, modern office interior, window in background, shallow depth of field, natural daylight from the window, neutral color palette” -
Use fixed patterns
Here are some templates you can reuse and tweak.
Portrait:
“RAW photo of a [age] year old [ethnicity] [man/woman], neutral expression, looking at the camera, studio portrait, 85mm lens, f1.8, shallow depth of field, soft diffused lighting, detailed skin texture, natural skin tones, 4k”
Full body:
“full body photo of a [age] year old [ethnicity] [man/woman], standing on a [description of ground], [type of outfit], shot on 35mm lens, natural daylight, realistic shadows, realistic proportions, background slightly blurred, 4k”
Product shot:
“studio photo of a [object], on a plain white background, softbox lighting from both sides, no reflections on lens, sharp focus, ISO 100, high detail, no text, no logo except [brand logo if needed]”
-
Use negative prompts aggressively
Most models respond well to clear negatives. Common ones:
“extra fingers, extra limbs, deformed hands, deformed feet, distorted face, asymmetrical eyes, ugly, low detail, lowres, low resolution, blurry, out of focus, noisy, overexposed, underexposed, watermark, text, logo, border, frame, nsfw, glitch, duplicate face, duplicate body” -
Match your model
Stable Diffusion style models like:
“highly detailed, ultra realistic, RAW photo, 8k, sharp focus, HDR”
Midjourney likes:
“high detail, realistic, photography, volumetric lighting, 8k uhd”
DALL·E likes concrete object and scene descriptions with fewer style buzzwords. -
Do small prompt edits, not huge changes
Generate 4 images.
Pick the best.
Edit prompt in small ways:
• “softer lighting”
• “slightly older face”
• “less saturated colors”
Do not rewrite the whole thing every time.
Example full prompt for a realistic scene:
“photo of a 30 year old Black woman sitting at a kitchen table, morning light from window on the right, holding a white coffee mug, casual grey t shirt, clean modern kitchen in background, shot on 35mm lens, f2.0, shallow depth of field, natural skin texture, subtle under eye shadows, realistic proportions, neutral color grading, 4k, high detail”
Negative:
“lowres, blurry, extra fingers, extra hands, duplicate face, distorted face, strange reflections, cartoon, painting, 3d, oversaturated colors, watermark, text, logo”
If you post one of your current prompts and a sample output, people can help tune the wording pretty fast.
You’re already getting solid stuff from @sonhadordobosque. I’ll add some angles that helped me move from “generic stock” to “did a photographer shoot this?”
I don’t fully agree that negatives + camera terms are the whole game. They help, but the big jump for me came from:
1. Stop describing, start art directing
Bad:
“photo of a woman in a city street at night, realistic, 4k”
Better:
“candid nighttime street photo of a 26 year old Latina woman crossing a wet downtown street, light rain, neon shop signs reflecting in puddles, she holds a transparent umbrella, slightly motion blurred background, realistic traffic lights, pedestrians in the distance, subtle city haze, realistic proportions”
Notice:
- Verbs / actions: “crossing,” “holds”
- Environment doing work: rain, neon, puddles, haze
That’s what kills the “generic stock” look.
2. Lock in a visual story in one sentence
I usually write a 1‑sentence story, then convert it into a prompt.
Story:
“A tired doctor sits alone in a hospital break room at 3 AM, staring at a coffee cup under harsh fluorescent light.”
Prompt:
“photo of a 40 year old Asian male doctor in scrubs, sitting alone at a small table in a hospital break room at 3 AM, elbows on table, hands wrapped around a white coffee cup, subtle dark circles under eyes, harsh fluorescent ceiling lighting, slightly messy room with lockers and notice board in background, realistic color cast from fluorescents, natural pose, 4k”
You basically force the model into a scene that feels like a screenshot from real life.
3. Use imperfections on purpose
A lot of “too perfect” images scream AI. I actually add small flaws:
- “slightly frizzy hair”
- “subtle under eye bags”
- “mild skin blemishes”
- “slightly crooked smile”
- “a few wrinkles around eyes”
Example:
“RAW photo of a 32 year old white woman, casual hoodie, slightly messy hair, mild acne and visible pores, tired but warm smile, office background with cluttered desk, natural window light from left, realistic color balance, no skin smoothing, 4k”
That “no skin smoothing” bit can matter more than 3 extra “ultra detailed” buzzwords.
4. Control clothing & era like a wardrobe stylist
Instead of:
“man in suit, realistic photo”
Try:
“photo of a 45 year old Black man wearing a tailored navy business suit, white dress shirt, slim dark grey tie, top button closed, polished black oxford shoes, standing on a sidewalk in front of a modern glass office building, crisp but natural fabric texture, realistic wrinkles in clothing”
You can also set era:
- “late 90s casual clothing”
- “2010s tech startup hoodie and jeans”
- “1970s fashion, wide collar shirt, flared pants”
This pushes the model away from the same few generic outfits.
5. Dial in background with 2–3 concrete props
Generic:
“living room background”
Better:
“modern living room with a grey fabric sofa, small wooden coffee table with a plant, large window with sheer white curtains, warm floor lamp in corner”
Even if the model messes up small details, it feels like a place, not a backdrop.
6. Treat composition like actual photography
I slightly disagree with avoiding “cinematic” all the time. It can be fine if you give it structure. But instead of that vague word, try composition terms:
- “rule of thirds composition”
- “subject centered in frame”
- “subject on the right third, looking into empty space”
- “low angle shot”
- “over the shoulder shot”
Example:
“photo of a 30 year old Indian woman working on a laptop at a cafe table, over the shoulder shot from behind her left side, screen slightly visible, shallow depth of field, busy cafe background softly blurred, warm afternoon window light, 4k”
Composition terms help the model decide what is actually in frame.
7. Make a reusable “realism shell” and swap the middle
I keep a base structure and only change subject + setting. Like:
“RAW photo, realistic proportions, no skin smoothing, sharp but natural focus, subtle film grain, natural color grading, no HDR look, 4k resolution”
Then:
Portrait:
“RAW photo of a 29 year old Middle Eastern man, short beard, casual dark green t‑shirt, sitting on a park bench, trees softly blurred in background, overcast daylight, realistic proportions, no skin smoothing, subtle film grain, natural color grading, 4k resolution”
Street:
“RAW photo of an elderly white woman walking her small dog on a quiet suburban street in the morning, light fog, parked cars on the side, wet asphalt, realistic proportions, subtle film grain, natural color grading, 4k resolution”
Same shell, different story.
8. Prompt what to keep small, messy, or out of focus
Also helpful:
- “background elements slightly out of focus”
- “cluttered desk but not the main focus”
- “people in background small and indistinct”
Example:
“photo of a 35 year old Brazilian man working late at a computer in a dark office, only monitor light on his face, messy desk with papers and a mug, background office chairs and shelves barely visible and out of focus, realistic screen glow on skin, muted colors, 4k”
That avoid the weird “everything is equally important” AI look.
9. One full example prompt you can dissect
Subject:
“photo of a 27 year old Black woman with short natural hair”
Prompt:
“photo of a 27 year old Black woman with short natural hair, sitting at a small round table in a cozy cafe, late afternoon, wearing a light brown sweater and simple gold earrings, hands wrapped around a ceramic coffee mug, soft window light from the left side, warm practical lights visible in the background, shallow depth of field, background patrons softly blurred, realistic skin texture with subtle pores and minor blemishes, natural expression, slightly thoughtful, realistic proportions, natural color grading, subtle film grain, 4k”
Negative (short so it does not fight everything):
“cartoon, painting, illustration, 3d render, extra fingers, deformed hands, distorted face, lowres, oversharpened, heavy makeup, plastic skin, watermark, text, logos”
If you post one of your “generic” prompts plus the kind of image you wish it was making, people can usually tweak like 6–8 words and suddenly it looks way more real. The trick is less “magic keywords” and more “be the art director who knows where the light is, what the person is wearing, and what tiny flaws make it feel human.”
Skip the magic-word hunting. Think in systems instead, especially if you’re hopping between Midjourney, DALL·E, SD, etc. I’ll add stuff that plays well with what @sonhadordobosque wrote, but from a different angle.
1. Build a “prompt scaffold” instead of freestyling
A lot of generic output comes from mixing style, subject and tech terms randomly. Use a fixed order so your brain and the model know what matters first.
Example scaffold:
- Subject
- Action / emotion
- Setting / time of day
- Camera / lens / composition
- Lighting logic
- Surface & material details
- Post‑processing style
- Constraints (realism, no stylization, etc.)
- Very short negative prompt
One workable template:
[SUBJECT + AGE + ETHNICITY], [ACTION / EMOTION], [SETTING + TIME], [CAMERA + LENS + ANGLE], [LIGHTING], [MATERIAL / SURFACES], [POST LOOK], [REALISM CONSTRAINTS], [NEGATIVES]
Example:
“35 year old East Asian woman software engineer typing on a laptop at her kitchen table at 11 pm, apartment interior, 50mm lens, eye level, subject in right third of frame, warm overhead pendant light and cool laptop screen light mixing on her face, wooden table with visible grain, ceramic mug with tea stains, slight noise like high ISO, realistic skin texture, no makeup, no HDR, no cinematic color grading, cartoon and illustration and 3d render and text and watermark removed”
Notice I did not stack 20 buzzwords like “hyper realistic ultra 8k.” Fewer, clearer instructions.
2. Pick one realism anchor: a physical constraint
Where I slightly disagree with “just art direct the scene” is that even a well directed scene can still feel AI if the physics are off. You want one specific constraint that tells the model “this must obey reality.”
Examples of realism anchors:
- A specific lens + f‑stop
- “35mm lens at f/1.8, shallow depth of field, background blur but eyes perfectly in focus”
- Lighting tied to a believable source
- “only light source is a single desk lamp to the left”
- Camera limitations
- “slight ISO noise in darker areas”
- “subtle motion blur on fast moving hand, face sharp”
- Color constraints
- “neutral white balance, no teal and orange, no stylized color grading”
You do not need all of those at once. One or two is enough to pull it away from ‘AI gloss.’
Prompt using a couple of anchors:
“candid photo of a man jogging on a city sidewalk at dawn, 35mm lens at f/2.2, slight motion blur in legs while face remains sharp, cool blue ambient light with warm streetlights in distance, visible ISO noise in shadows, neutral color balance, realistic proportions, no skin smoothing, no HDR”
3. Structure emotions, not adjectives
Instead of piling on “sad, dramatic, realistic,” define what is causing the emotion and how it shows physically.
Bad:
“sad woman, realistic, emotional, photography”
Better:
“young woman in her late twenties sitting on a closed toilet in a small bathroom, phone in her hands with unread message on screen, shoulders slightly slumped, eyes slightly red, trying not to cry, small overhead light, cold tile floor, muted colors, realistic proportions”
The emotion comes from posture, props, and light, not the word “sad.”
Try this mini‑pattern for human scenes:
[emotion trigger] + [body reaction] + [face reaction] + [environment reflecting mood]
4. Use context pairs instead of long lists
A lot of people dump long chains like:
“street, city, bokeh, cinematic, moody, depth of field, filmic, realistic”
The model sometimes latches onto just 1 or 2 and forgets the rest.
Pair related concepts tightly:
- “overcast sky, diffuse shadow”
- “office fluorescent tubes, slightly green color cast”
- “late afternoon sun, long soft shadows”
- “old digital camera, mild sensor noise”
Putting them as pairs hints at causality, which realism loves.
Example:
“documentary style photo of a middle aged man repairing a bicycle in a small garage, single overhead fluorescent tube with slight green color cast, cluttered shelves in shadow, concrete floor with oil stains, 35mm lens, natural depth of field, neutral color grading, realistic proportions”
5. Calibrate style strength instead of yelling “realistic” 10 times
Different generators respond to vague words like “photorealistic” in different ways. Rather than repeating it, pin “realism” to a visual reference or a style tier.
Patterns that often work better:
- “shot on a modern DSLR, clean color, no film emulation”
- or “shot on 35mm Portra 400, subtle grain, soft contrast, warm tones”
- or “ungraded RAW photo, flat contrast, slightly desaturated”
Pick one. If you want clinical realism, something like:
“straight out of camera RAW look, flat contrast, neutral colors, no dramatic lighting, no film grain”
For “stunning but still real”:
“subtle commercial photography look, gentle contrast boost, clean whites, no heavy vignetting, no split toning”
You can even think in “tiers”: documentary, commercial, lifestyle, fashion, etc.
6. Use relative composition, not just photography jargon
I agree composition words help, but a lot of models overinterpret “rule of thirds” or “wide shot.” They sometimes respond better to relative instructions.
Examples:
- “subject fills about one third of the frame height”
- “background buildings taller than subject but not cropped”
- “head and shoulders only, no hands visible”
- “full body in frame, feet not cropped”
If you keep getting weird crops or floating torsos, try:
“full body clearly visible from head to shoes, centered in frame, feet touching the bottom edge of the frame, enough headroom above”
That kind of explicit framing often fixes the uncanny crop issues.
7. Try iterative prompting instead of a single perfect wall of text
This is where I diverge from the “perfect prompt” thinking. A lot of amazing images you see are two or three prompt passes.
Workflow example:
-
Pass 1: Simple, structural prompt
- Aim: get pose, composition, and lighting roughly right.
- “photo of a 30 year old man sitting on a train by the window, 50mm lens, side view, early morning light through window, subtle reflections on glass, neutral color grading, realistic proportions”
-
Pass 2: Refine with extra details or an image prompt
- Feed the best result back in with:
- “same scene, slightly more detail on clothing fabric and window reflections, subtle film grain, no skin smoothing, no HDR”
-
Pass 3: Correct a specific flaw only
- “keep composition and pose, correct hands and fingers, remove any distortions in eyes, preserve natural skin texture”
You do not need to say everything in one monster paragraph. Constrain what changes on each pass.
8. Controlled negatives instead of blacklist novels
You already saw a clean negative from @sonhadordobosque. I would double down on keeping negatives task specific.
-
Portraits:
- “cartoon, painting, illustration, 3d render, extra fingers, extra limbs, distorted face, plastic skin, exaggerated makeup, watermark, text”
-
Products:
- “lowres, heavy reflections, warped labels, melted edges, blurry text, extra logos, watermark, cartoon, 3d render”
-
Architecture:
- “impossible geometry, melted buildings, repeating windows, warped doors, cartoon, illustration”
Avoid giant catch-all negatives like “blurry, out of focus” when you actually want shallow DOF. Instead, say:
“sharp focus on eyes, background may be out of focus but not distorted”
9. When to ignore realism on purpose
Ironically, sometimes images look more real if you allow a tiny bit of stylization rather than forcing brutal realism.
Some subtle stylization flags that still read as photo:
- “subtle lifestyle magazine look”
- “gentle warm tone curve like a phone camera filter”
- “soft vignette like a cheap lens”
Use that if your model’s default photo style feels sterile. It gives the illusion of a human photographer’s taste.
10. Quick prompt patterns you can steal
Corporate headshot
“corporate headshot photo of a 45 year old Black woman executive, standing in front of a softly blurred modern office background, 85mm lens at f/2, shoulders and head in frame, natural expression, slight smile, simple navy blazer and white blouse, straight out of camera look with neutral color balance, realistic skin texture with pores and fine lines, no skin smoothing, no harsh contrast, no cinematic color grading, cartoon and illustration and 3d render and watermark removed”
Everyday indoor moment
“photo of a middle aged white man washing dishes in a small apartment kitchen at night, overhead warm ceiling light, small window showing dark city outside, stainless steel sink with water droplets, ceramic plates with minor chips, 35mm lens, medium shot from behind and slightly to the side, natural clutter on counter, neutral color grade, realistic proportions, no beauty retouching, no HDR”
@sonhadordobosque is really strong on scene direction and character detail, which pairs nicely with this more “technical constraints” angle. Try merging both: define the micro-story like they showed, then wrap it in a scaffold with specific physics, camera behavior, and a small, targeted negative prompt.
If you post one of your generic prompts and the exact type of photo you’re chasing (e.g. “LinkedIn headshot,” “candid street photo,” “product on white”), people can help you rewrite it in that scaffold and you’ll see pretty quickly which part was holding you back.