Write photorealistic AI image and video prompts with lens, light, material and motion detail — plus worked examples, decision criteria and fixes.
Photorealism in AI generation is rarely a style problem. It is a specification problem. Hand two engines the same sentence — a woman in a cafe, photorealistic, 8k — and you get the same averaged result: smooth skin, soft haze, a background that dissolves into mush. Hand them a lens, an aperture, a light direction, and a material description, and the output changes category entirely.
That gap is what this guide covers. Not magic words and not a hidden parameter, but the habit of writing prompts the way a photographer writes a shot list — with constraints specific enough that the model has nothing left to invent.
Everything below applies to stills built in the AI image generator and to shots animated in the AI video generator, because the same handful of variables decides whether a frame reads as a photograph or as a render.
Realism is a specification problem, not a style
Most weak photorealistic prompts ask for a quality instead of describing a cause. Hyperrealistic, ultra-detailed, 8K, masterpiece, best quality — these are outcome claims. A model cannot act on an outcome claim, so it guesses, and the guess is the over-smoothed average that viewers instantly flag as synthetic.
A strong prompt instead names the physical conditions that produce realism in a real capture: which lens was used and how much of the scene is in focus, where the light comes from and how hard it is, what the surfaces in frame are made of and how they respond to that light, and what the camera would have frozen or blurred at that moment. When those four things agree with each other, realism appears as a side effect. When they disagree — soft window light plus hard noon sun plus a ring-light catchlight — the model splits the difference into something vague and waxy.
Here is the translation table worth keeping next to your editor:
| Vague claim | What it hides | What to write instead |
|---|---|---|
| photorealistic | everything | 85mm at f/2, overcast window light camera right |
| 8K, ultra-detailed | resolution, not detail | pore-level skin texture, visible linen weave |
| cinematic lighting | mood | one large softbox at 45 degrees camera left, 4:1 ratio |
| natural light | quality and direction | north-facing window two meters from subject, white wall fill |
| blurred background | the cause of the blur | 85mm, f/1.8, subject 1.5 m from lens, background 6 m behind |
| studio setup | geometry | key at 45 degrees, black card negative fill opposite, rim light behind |
The practical test before you generate: could a real photographer actually shoot this frame as described, with gear that exists? If the answer is no, the model will improvise, and improvisation is where artifacts live. If the answer is yes, you have a spec rather than a wish.
Anatomy of a prompt that renders like a photograph
A reliable photorealistic prompt has six layers. They do not need a rigid order, but each one should be present or deliberately dropped.
Subject and state. Be concrete about who is in frame, what they are doing, and what condition they are in. Age range, build, clothing fabric, posture, and gaze direction do more work than facial adjectives. Instead of "beautiful woman," write "woman in her early thirties, fine lines at the eyes, hair pulled back with loose strands at the temples, leaning forward with her weight on one elbow." Specificity of state is what makes a subject look photographed rather than rendered: lips slightly parted mid-word, a crease across the shoulder where the shirt turns, one hand out of focus in the foreground.
Environment and surfaces. Name the place and its material reality. "A cafe" is not enough. "A narrow cafe with a zinc countertop, tiled backsplash, and condensation on the window" works because it gives the model surfaces to light and bounce. Context also anchors the palette: overcast northern daylight, sodium streetlamps, fluorescent ceiling panels.
Lens and distance. Focal length plus distance plus camera height determine perspective, and perspective is the strongest realism cue available. A 24mm close to a face stretches the nose; an 85mm at f/1.8 compresses features and melts the background. Getting this wrong is the most common reason an otherwise decent image feels off without anyone being able to say why.
Light. Direction, quality, color, ratio. Not "moody lighting," but "single softbox at 45 degrees camera left, warm practical lamp behind the subject, deep shadow on the right side of the face."
Material micro-detail. Skin texture, fabric weave, metal wear, glass refraction, dust, fingerprints, scuffs. Realism lives mostly in micro-detail at the correct scale — not in more detail everywhere.
Output intent. Aspect ratio, framing, and whether the frame should read as editorial photography, a film still, a product shot, or documentary reportage. This shapes grading and composition more than most people expect.
Order matters too. Front-load subject and action, then camera, then light, then material, then output intent. Short clause stacks separated by commas behave more predictably than long grammatical sentences, because they read like a spec sheet rather than a story.
Camera language that actually changes the output
Models respond to photographic vocabulary in fairly predictable ways once you use it precisely.
Aperture controls separation. f/1.4 to f/2 gives a thin plane of focus with creamy falloff; f/5.6 keeps a person sharp with a soft room behind; f/8 keeps an interior readable; f/16 adds the faint diffraction softness of a landscape lens. Naming the aperture beats writing "blurred background," because it tells the model how fast the falloff should be.
Focal length controls perspective. 24–35mm for environmental context and slight edge stretch, 50mm for a neutral human view, 85–135mm for compression and flattering portraits, 200mm for stacked layers in a street scene.
Shutter controls motion. A fast shutter freezes a splash mid-air; a slow shutter turns passing headlights into streaks. In video prompts, shutter behavior becomes motion blur, and inconsistent motion blur is one of the fastest ways to make a generated clip feel synthetic.
Film stock and grain add texture. Naming a stock type — fine-grain daylight negative, pushed high-ISO color negative, black-and-white stock with visible grain — changes contrast, highlight roll-off, and color response far more than the word "cinematic" does.
Sensor and format change rendering character. Full-frame with a fast prime looks different from medium format with deep tonality, and both look different from a phone camera with computational sharpening. Choose deliberately: phone-camera realism is sometimes exactly what a social-first piece needs, and medium-format polish is sometimes exactly wrong.
Camera height is the most ignored variable. Chest height reads neutral, hip height reads heroic, floor height reads documentary, and overhead reads graphic. Two prompts with identical text but different stated heights produce genuinely different compositions.
Lighting: describe geometry, quality, and ratio
Lighting is where prompts most often fail, because people describe the feeling of light instead of its geometry.
Start with direction. Side light sculpts, front light flattens, backlight separates, and a 45-degree key with negative fill is the workhorse of commercial photography. Then quality: a small source produces crisp shadow edges and hot specular highlights; a large diffused source wraps and reduces contrast. Then color, ideally in Kelvin — 3200K tungsten warmth, 5600K daylight neutrality, 7000K overcast blue. Then ratio, which is the most underused control in photorealistic prompting: 4:1 key-to-fill reads dramatic but natural, 1:1 reads flat catalog photography, 8:1 reads noir.
Setups that consistently produce believable frames:
- Window portrait. One large north-facing window as key, white wall as fill, no other source, subject two meters from the glass, faint shadow falloff on the far cheek.
- Product on location. Key softbox at 45 degrees, black card for negative fill opposite, thin rim from a nearby practical, visible brushing on the metal, dust at the base.
- Night street. Single sodium overhead, wet asphalt bounce, deep shadow falloff, neon practical three meters behind the subject for edge separation.
- Interior documentary. Mixed daylight and overhead fluorescents, slight green cast, uneven exposure across the room, one blown highlight on a window frame.
One lighting logic per image. If you describe two contradictory setups, the render averages them into a look that belongs to neither, and that averaged look is the visual signature of weak AI output.
Materials and micro-detail: where fidelity lives
At normal viewing size, viewers do not audit geometry — they audit surfaces. Skin, fabric, metal, glass, and grit carry nearly all of the realism signal.
Skin needs pores, fine vellus hair, subtle redness around the nose and knuckles, a slight oil sheen on the forehead, and uneven tone. Light bleeding through the edge of an ear or nostril is the difference between a face and a mannequin; when it is missing, no amount of sharpening recovers the image.
Fabric needs weave, drape, and weight. Linen creases sharply, wool absorbs light, satin produces a narrow specular band, denim shows twill and fading at the seams. Naming the fiber is more effective than naming the color.
Metal needs wear. Brushed aluminum has anisotropic highlights, brass tarnishes in crevices, and chrome reflects the room — which means the room has to be described for the reflection to make sense.
Glass needs refraction and imperfection: faint fingerprints, dust at the base, a visible horizon line through the liquid. Perfect glass reads as CGI immediately.
Environmental grit sells the frame — dust motes in a light beam, a scuff on the floor, crumbs on a table, condensation rings, a worn doorframe edge. Add one or two per image. More than that and the prompt becomes noise, and noise dilutes the details that were working.
From still frame to motion: keeping realism through the shot
Video raises the bar, because realism has to survive time. A frame that looks photographic can fall apart the moment it moves. Three things matter most.
Motion physics. Describe weight and inertia. A coat lags behind a turn, hair settles after the head stops, liquid splashes outward before it falls. Prompts that describe forces — wind from camera left, a heavy door swinging inward, fabric pulled taut by a raised arm — produce better movement than prompts that only describe appearance.
Camera behavior. Use real rig vocabulary: slow dolly in, handheld drift with slight breathing, gimbal walk at chest height, locked-off tripod, crane up. Each implies a specific amount of shake and a specific parallax. Combining a locked-off description with handheld energy gives the model contradictory instructions, and the result usually looks like a floating camera with no operator.
Temporal consistency. Busy scenes with many small elements drift over time. Fewer subjects, simpler backgrounds, and one clear focal point hold together across a clip. When a shot must be complex, generate a strong still first and animate from it so composition is locked and the model only has to solve movement.
The workflow that saves the most time: build the key frame as an image, refine it until it passes your realism checklist, then write a motion prompt containing only camera and subject movement. Repeating the whole lighting and material spec in the video prompt often causes the model to reinterpret the scene instead of moving a camera through it — and reinterpretation means the face changes, the light jumps, and the wardrobe drifts.
Three worked prompts and an iteration loop
Portrait. Weak: portrait of a woman in a cafe, photorealistic, 8k. Stronger: woman in her early thirties seated at a zinc cafe counter, elbows on the counter, looking just past the lens, 85mm at f/2, overcast window light from camera right, white wall fill, fine-grain daylight negative, full-frame, shallow depth of field, slight grain, 3:2. The second version works because every clause is shootable: a lens, an aperture, a light direction, a fill source, a stock, and a format.
Product. Weak: luxury watch on a table, studio lighting, hyperrealistic. Stronger: stainless steel chronograph on grey concrete, 100mm macro at f/11, large softbox at 45 degrees camera left, black card negative fill on the right, thin rim light from behind, visible brushing on the case, micro-scratches on the clasp, dust at the base of the crystal, subtle reflection in the concrete, 4:5.
Street. Weak: rainy city street at night. Stronger: narrow Tokyo side street after rain, one sodium streetlamp overhead, wet asphalt reflecting signage, steam from a vent at frame left, 35mm at f/2.8, handheld height, deep shadow on the right wall, high-ISO color negative with visible grain, 16:9.
Then iterate like a technician rather than a gambler. Change one variable per round — the aperture stays while the light moves; the light stays while the lens changes. Generate three or four variations per prompt version; if none resolves the issue, the problem is in the prompt, not the seed. Log the prompt, seed, and settings next to the output, because photorealism is cumulative and you will not remember which clause produced the good frame.
A quick debugging table helps more than intuition:
| Symptom | Likely cause | Fix |
|---|---|---|
| Plastic, waxy skin | flat front light, no micro-detail | add side key, specify pores and fine hair, remove smoothing words |
| Everything sharp | no aperture or focal length | state 85mm at f/2 and subject distance |
| Muddy, directionless shadows | two conflicting keys | keep one key, add negative fill |
| Chaotic composition | too many subjects | cut a subject before adding detail |
| CG-looking glass or metal | no wear, no reflection context | add fingerprints, tarnish, describe the reflected room |
| Uncanny motion in video | no force described, mixed rig language | name a single rig behavior and one physical force |
Decision criteria, common mistakes, and a pre-publish checklist
No single engine wins everywhere, so match the tool to the job and keep a small set of tested options. Criteria worth writing down before you commit: subject type (portraits, products, architecture, and food reward different strengths), texture fidelity (some models handle skin well and repeated patterns badly), text rendering (test signage early if words must be readable), consistency across shots (sequences favor reference-image and image-to-video workflows), iteration speed (fast drafts matter more than peak quality while you are still exploring composition), motion quality (stable cameras and strong subject motion rarely come from the same setting), and aspect-ratio support.
Run a fixed test across candidates: same prompt, three seeds, two aspect ratios, compared at a 100 percent crop rather than thumbnail size. Standardize a house template with slots for subject, lens, light, material, and format so results are comparable between people and revisits. If you want starting points rather than a blank page, browse the prompt library and the video templates, then adapt them to your own subject.
Mistakes that break realism most often:
- Stacking quality words. Ultra, hyper, 8K, and best quality together do not add fidelity; they add averaging.
- Conflicting light. Two keys, or hard sun with an unaccounted soft fill, produce mush.
- Lens mismatch. Wide-angle perspective described with telephoto compression reads as wrong even to untrained viewers.
- Plastic skin. Caused by missing pore-level detail and by front lighting that erases texture variation.
- Overloaded scenes. Each added subject takes attention away from the details that sell realism.
- Ignoring aspect ratio. A 2.39:1 crop changes composition; describing a vertical portrait inside a wide frame wastes the frame.
- Naming a camera brand. Naming a body rarely helps; naming format and lens behavior does.
- No negative fill. Without shadow, nothing looks three-dimensional.
Checklist before publishing or animating:
- Exactly one lighting logic, with stated direction and quality.
- A focal length and aperture, with depth of field that matches them.
- A subject with a physical state, not just an appearance.
- At least two materials described by how they react to light.
- One piece of environmental grit at the right scale.
- A palette that comes from the scene rather than a filter.
- An aspect ratio that fits where the image will be used.
- A setup a photographer would recognize as shootable.
- For video: rig language for the camera and weight for the subject's motion.
- A saved prompt and settings file for reuse.
FAQ
Do I need a long prompt to get photorealism? No. You need specific clauses, not many. A 35-word prompt with a lens, an aperture, one light source, and two material details usually beats a 150-word prompt full of quality words.
Which single addition improves realism most? Aperture plus focal length. Naming something like 85mm at f/2 fixes perspective and focus falloff at once, and those are the two cues viewers use most when judging whether an image is a photograph.
Why does my subject look waxy? Usually flat front lighting plus missing micro-detail. Add a side key, specify pores and fine hair, and remove smoothing adjectives such as flawless or airbrushed.
Can the same prompt work for stills and video? The descriptive half transfers, but the motion half must be rewritten. For video, state camera rig behavior and physical forces, and drop redundant lighting detail when you animate from a locked reference frame.
How do I keep a character consistent across shots? Lock a reference image, keep wardrobe and lens language identical, vary only camera angle and action, and note the settings that worked. Consistency is workflow discipline more than a prompt trick.
How many variations before changing the prompt? Three or four per prompt version. If none resolves the issue, the prompt is the problem, not the seed. Change one variable, then re-test.
Does post-processing fix a weak prompt? Only marginally. Grade, grain, and grain-matching can unify a set, but they cannot invent subsurface light in skin or a believable shadow that the render never produced.
Build the shot, then generate it
The shift from prompting for style to prompting for conditions is what separates unpredictable luck from a repeatable process. Describe a lens, a light, a surface, and a format; keep one lighting logic; iterate one variable at a time; log what worked. That discipline travels across subject matter and across engines, which is exactly why it is worth learning once.
When you are ready to put it into practice, start with a still in Orelon, refine it against the checklist above, then bring it into motion with the AI video generator so the camera moves through a frame you already trust. If you are weighing engines for a specific look, the alternatives hub breaks down where each one is strong, and the Orelon blog goes deeper on workflow. Cinematic ideas in motion begin with a frame that already looks real.