What gives an AI image away is usually not missing detail but a physics error: the light comes from one place while the shadow falls to another, the skin is impossibly smooth, the background is sharper than it should be.
Photorealism comes from writing that physics into the prompt. These models were trained on photographic archives whose images carried technical metadata; speak that language and they answer much better.
The four techniques below account for most of what moves an image towards photography.
1. Describe the lens
Focal length fundamentally determines how a scene looks, and writing it into the prompt works directly.
85mm gives natural facial proportions in portraiture and softens the background. 50mm is closest to human vision and looks natural in everyday scenes. 35mm gives a wider frame that shows a person with their environment — the documentary standard.
Aperture can be written too: f/1.8 blurs the background markedly, f/8 keeps everything sharp. "Shallow depth of field" does the same job more generally.
Wide-angle descriptions (16mm, 24mm) distort faces in portraiture but are the right choice for architecture and interiors.
2. State the light's direction and quality
Light is the single detail that makes an image convincing or fake, and it is the most frequently omitted information in prompts.
Write the direction: "soft window light from the left", "a single hard source from above", "rim light from behind". Write the quality: soft light spreads shadows and is preferred in portraiture; hard light creates sharp shadows and reads dramatic.
"Overcast daylight" gives soft, shadowless illumination and is the most reliable light description for fashion and architecture. "Golden hour" is warm and low-angled, and used as backlight it creates a glow along the hair.
If you are placing a product into a scene, light direction is doubly critical: if it disagrees with the shadow in your own product photo, the image reads as fake.
3. Stop the skin looking plastic
Models over-smooth skin by default; the result is a technically flawless face that does not read as human.
The phrases that fix it are simple: "natural skin texture", "visible skin pores", "unretouched skin", "fine lines". Even one of them makes a visible difference.
The second helper is describing soft rather than hard light. Soft light makes skin texture more believable; hard light emphasises the smoothing.
Specifying age and character helps too. Instead of "a woman", write "a woman in her fifties with pronounced smile lines" — that pulls the model away from its default flawless-young-face tendency.
4. Choose the right tool
Prompt technique takes you a certain distance; beyond that, tool choice decides.
Flux leads clearly on photorealism. Hands and faces — long the two weakest points — come out consistent here. If the work has to look like a photograph, this is the right starting point.
Stable Diffusion's photoreal checkpoints also produce excellent results, but finding and tuning the right model takes effort.
Midjourney always stays slightly stylised: the light a little more dramatic, the colours a little more harmonious, the composition a little more staged. An advantage in artistic work, a flaw when you need a photograph.
Frequently asked questions
Does describing a lens really make a difference?
Yes. Models were trained on photographic archives where most images carried lens metadata. "85mm" naturalises facial proportions, "35mm" gives a wider frame including the environment, "f/1.8" blurs the background.
Why does the skin look plastic?
Models over-smooth skin by default. Adding "natural skin texture" or "visible skin pores" corrects it markedly. Describing soft rather than hard light also makes the texture more believable.
Which tool for photorealism?
Flux leads clearly; hands and faces come out consistent. Stable Diffusion's photoreal checkpoints are also excellent but need effort to find and tune. Midjourney always stays slightly stylised.