The gap between an AI image that looks fake and one that fools people is almost never the subject. It’s the camera language. Photorealistic AI art prompts work because they tell the model to imitate a real photograph — a real lens, a real aperture, real light falling a real way. Leave that out and you get the plastic, over-smooth “AI face” everyone clocks in half a second. Put it in and skin gets pores, backgrounds melt into honest bokeh, and the whole frame reads like it came off a memory card.

I’ll walk through the four things that actually move realism — lens, aperture, lighting, and film stock — then hand you full prompts you can paste into Midjourney or Stable Diffusion right now. No theory you can’t use.

Why photorealistic AI art prompts need camera language #

Here’s the mental model. The model has seen millions of images tagged with camera data: “85mm,” “f/1.4,” “shot on Kodak Portra.” Those tags cluster with a specific look — shallow focus, flattering compression, fine grain. When you write those words, you’re pointing the model straight at that cluster. Skip them and it defaults to a vague, well-lit, slightly-too-clean average that screams render.

So a photoreal prompt is really a fake photo caption. You’re not describing a painting. You’re describing the gear and the conditions a photographer would’ve used. Once that clicks, the keywords stop feeling random and start feeling like a checklist.

Lens and focal length (the realism backbone) #

Focal length changes the whole feel of a shot, and the model knows it. Match the lens to the subject:

  • 85mm — the portrait standard. Flattering compression, natural proportions. Default for faces.
  • 50mm — “normal” lens, close to how the eye sees. Great for environmental portraits and street.
  • 35mm — wider, keeps context. Documentary and reportage feel.
  • 24mm wide-angle — big spaces, slight edge stretch. Interiors, landscapes, drama.
  • 100mm macro — tight detail, insects, eyes, textures.
  • 200mm telephoto — heavy background compression, wildlife, candid distance.

Pair the focal length with a real camera body if you want extra grip — a full-frame mirrorless body, a medium-format back. You don’t need to name a brand to get the effect; “full-frame DSLR” or “medium format” already steers it.

Aperture and depth of field #

Aperture controls how much of the frame is sharp, and it’s the fastest way to kill the flat AI look. Wide apertures throw the background out and isolate the subject:

  • f/1.4 – f/1.8 — creamy bokeh, razor-thin focus, dreamy portraits
  • f/2.8 — soft background, subject sharp, a touch more context
  • f/5.6 – f/8 — balanced, most of the scene sharp, good for groups
  • f/11 – f/16 — deep focus front to back, landscapes and architecture

In Stable Diffusion you can lean on the effect with weighted tokens like (shallow depth of field:1.2) and (bokeh:1.1) alongside the f-stop. Belt and suspenders, but it works.

Midjourney:
candid portrait of an older fisherman, weathered face, 85mm, f/1.8, shallow depth of field, soft window light, full-frame --ar 4:5 --style raw

Stable Diffusion:
candid portrait of an older fisherman, weathered skin, (skin pores:1.1), 85mm lens, f/1.8, (shallow depth of field:1.2), bokeh, soft window light
Negative: plastic skin, smooth, airbrushed, cartoon, oversaturated, distorted hands, extra fingers

Lighting that sells the realism #

Lighting is paramount. Real photos have a direction and a quality of light; AI defaults to flat, sourceless illumination that reads as fake. Name the light:

  • Golden hour — warm, low, long shadows. Forgiving and flattering.
  • Soft window light — gentle directional, natural falloff. Portrait gold.
  • Overcast / soft natural light — even, low-contrast, true skin tones.
  • Rim lighting — bright edge that separates subject from background.
  • Hard directional sun — crisp shadows, high contrast, midday energy.
  • Studio softbox / Rembrandt lighting — controlled, that little triangle of light under the eye.
  • Blue hour — cool, even, post-sunset glow for moody exteriors.

One light direction beats a pile of light adjectives. “Soft window light from the left” outperforms “beautiful dramatic cinematic lighting” every time, because it gives the model a physical setup to solve.

Midjourney:
a chef plating a dish in a dim restaurant kitchen, 35mm, f/2.8, rim lighting from behind, warm tungsten glow, documentary --ar 3:2 --style raw

Stable Diffusion:
a chef plating food in a dim kitchen, 35mm lens, f/2.8, rim lighting, warm tungsten, (realistic skin:1.1), film grain, documentary photo
Negative: flat lighting, studio, clean, cartoon, plastic, watermark

Film stock and color grade #

Film-stock names are a cheat code for color and grain. Each stock carries its own palette and texture, and the model has absorbed those associations:

  • Kodak Portra 400 — warm, soft skin tones, the modern portrait look
  • Kodak Gold 200 — nostalgic, golden, slightly punchy
  • Fuji Pro 400H — cool, pastel greens, airy
  • CineStill 800T — tungsten-balanced, neon halation, night-city mood
  • Ilford HP5 — classic black-and-white grain and contrast
  • Polaroid / instant film — soft, washed, square, vintage

Add the stock at the end of the prompt like a real photographer noting it in a caption: “shot on Kodak Portra 400.” Keep it to one stock — mixing them muddies the grade.

Midjourney:
young couple laughing on a city street at dusk, 50mm, f/2.0, CineStill 800T, neon signs, halation, night --ar 3:2 --style raw

Stable Diffusion:
young couple laughing on a city street at dusk, 50mm lens, f/2.0, shot on CineStill 800T, neon glow, halation, film grain, photorealistic
Negative: oversharpened, plastic skin, cartoon, daylight, clean digital, deformed

The negative prompt does half the work #

In Stable Diffusion, a tuned negative prompt is non-negotiable for photoreal. It’s where you exile the exact tells that give AI away. Keep a default block and reuse it:

Negative (reusable photoreal block):
plastic skin, smooth skin, airbrushed, cartoon, illustration, 3d render, oversaturated,
overexposed, distorted hands, extra fingers, fused fingers, watermark, text, logo,
low quality, deformed, mutated, doll-like, waxy

Midjourney doesn’t take a full negative block, but --no handles the worst offenders — --no cartoon, plastic — and --style raw pulls back the house aesthetic so the photo reads straighter.

A repeatable photoreal recipe #

Put it on rails. Every time, in this order: subject and a real detail, focal length, aperture, lighting direction, film stock. That’s the spine of every prompt above. Fill the five slots, generate four, then fix the weakest slot — grain wrong, change the stock; too flat, change the light; too sharp and fake, drop the aperture wider and add skin-texture tokens.

Do that a dozen times and the checklist becomes muscle memory. You’ll stop writing “make it realistic” — which does nothing — and start writing the actual photo. If you want a head start, the brand’s generator scaffolds these camera-and-lighting prompts so you’re filling slots instead of staring at a blank box.