Faces are the hardest thing to make photorealistic in AI-generated imagery — and the gap matters more than it does for any other subject.
Humans are forensic when it comes to reading faces. Decades of research confirm that we can detect something off about a face in milliseconds, long before we can articulate what. We evolved to be. That makes AI portrait generation a uniquely demanding problem: it is not enough for the image to look good. It has to look real at the level of eye catchlights, skin pore distribution, and the natural asymmetry that every real face carries.
This post is a practical teardown of exactly where AI faces fail and what to do about each failure — covering the full workflow from prompting through post-processing. If you want the wider context on making all AI images look photographic, start with the make AI images look real pillar guide. Here we go narrow and deep on faces.
1.Why AI Faces Still Give Themselves Away
The short answer: AI models learned from retouched photography.
Most of the large-scale training data behind Midjourney, DALL-E, Stable Diffusion, and Flux includes billions of images that passed through beauty filters, Lightroom presets, or professional retouching before they were published. The model learns that 'a good portrait' means smooth skin, even lighting, and symmetrical features — because that is what the training distribution shows it most often.
The result is a portrait that is technically polished but reads as synthetic. It is not that the image has obvious artefacts. It is that it is statistically too clean. Real faces do not look like that. Real faces have pores. Real faces have one eye fractionally lower than the other. Real faces have a stray hair at the hairline and a slight shadow under the jaw that does not behave quite the same on each side.
The research reflects this: a 2026 UNSW study found that people are systematically overconfident in their ability to identify AI faces — but the tells they rely on are the wrong tells. The faces that fool people are not doing so through technical sophistication alone. They fool people because they look like the idealised portraits humans already aspire to produce. That cuts both ways: it means the worst AI faces can be improved substantially by adding back the imperfections that real photography never scrubbed out.
See also: why AI images get flagged for a breakdown of the structural reasons generation falls short across all image types.
2.The Face Realism Diagnostic Table
Before you can fix the problem, you have to identify which layer is breaking. Here is the complete diagnostic for AI portrait failure — each tell, the mechanism behind it, and the specific fix.
| Face Tell | Why It Looks AI-Generated | The Fix |
|---|---|---|
| Plastic, waxy skin | Model trained on retouched photos; averages out pore texture | Prompt: 'realistic skin pores, natural subsurface scattering, slight oiliness' |
| No catchlights in eyes | Model did not receive a physical light source to reflect | Specify light source and direction explicitly in prompt |
| Eyes too symmetrical / iris uniform | Real irises have radial pigment variation and limbal rings | Prompt: 'detailed iris texture, limbal ring, natural iris pigmentation' + portrait upscale |
| Over-perfect facial symmetry | Model optimises toward statistical average of training faces | Prompt: 'natural facial asymmetry, slightly uneven features' |
| Hair looks painted / uniform volume | No strand-to-strand variation; model lacks micro-hair data | Prompt: 'individual hair strands, natural flyaways, directional hair sheen' + side lighting |
| Flat, sourceless lighting | Model defaults to neutral frontal light when no source is given | Always name a specific light: 'soft window light from camera left at 45 degrees' |
| Expression feels posed or frozen | Model averages over all expressions, produces statistical mean | Prompt: 'candid mid-expression, natural smile, mid-laugh, thoughtful glance' |
| Background too clean | Perfect bokeh with no organic variation in out-of-focus areas | Specify lens and aperture to activate optical bokeh rendering |
| Skin tone flat across face | No subsurface scattering variation between cheek, nose, jaw | Prompt: 'natural skin tone variation, slight redness at nose and cheeks' |
3.Fixing Eyes and Catchlights
Eyes are where realism lives or dies in a portrait. A viewer does not consciously process a catchlight — but when it is absent or misplaced, the face reads as dead. Here is why: the cornea is a highly reflective curved surface. In every real photograph, that surface mirrors the light sources in the scene. A window produces a rectangular highlight in the upper iris. A ring flash produces a circular one at centre. Outdoor diffuse light produces an irregular bright zone.
When you give an AI model no light source, it has no physical reference for where to put that reflection. The result is an eye with flat, uniform iris coloring and a cornea that reflects nothing. It reads as painted.
The fix is simple but must be explicit. In your prompt, name a specific light source before anything else: 'soft natural window light from camera left.' That one instruction propagates through the entire image — it tells the model where shadows fall, where the skin lightens, and where the specular reflection in the eye should appear. The catchlight is a consequence of consistent lighting physics, not a feature to be prompted separately.
After generation, zoom into the iris at 100%. Real irises have radial pigment variation — darker at the centre, with visible crypts and collarette texture. AI irises tend toward uniform flat colour. A realistic AI image generator with portrait-specific upscaling can recover this iris detail by restoring high-frequency information the base model compressed.
4.Skin Texture: Getting Pores Back In
Skin is the largest tell because it covers the most surface area. The default AI output is smooth in a way that reads as impossible — not young, not healthy, but specifically as if the skin has been texture-erased.
Two levels of fix:
At the prompt level: include 'realistic skin pores, natural subsurface scattering, slight skin oiliness, faint freckles, light laugh lines, natural tonal variation.' The phrase 'subsurface scattering' is particularly effective — it instructs the model to simulate the way light penetrates skin and scatters internally before reflecting back, which is the physical process that gives real skin its luminous quality rather than its plastic alternative.
Remove texture-killing terms: 'flawless skin,' 'porcelain complexion,' 'perfect smooth skin' are all direct instructions to erase pore texture. Remove them entirely.
At the post-processing level: use a portrait detail enhancer — Imagera's detail upscaler runs specifically on the face region and recovers micro-detail that compression removed. For skin specifically, this surfaces pore structure, fine facial hair, and the slight tonal shift between the nose wing and the cheek that makes skin look like it has zones rather than a uniform painted surface.
5.Asymmetry: The Signature of a Real Face
Human faces are not symmetrical. Both eyes are never exactly the same height. The jaw angles fractionally differently on each side. The corners of the mouth sit at subtly different positions. These are not flaws — they are what every viewer's visual system expects to find when looking at a face, even without consciously noticing.
AI models produce faces that are more symmetrical than any real person. The model has been trained to produce a statistical average of faces, and averages are symmetrical by definition. This is a systematic bias, not a random one, and it is one of the clearest structural signals of generation.
Prompt directly against it: 'natural facial asymmetry, slightly uneven features, natural face structure.' Then in post-processing, examine the face at the horizontal midline. If both halves mirror precisely, use inpainting on one eye or jaw region to introduce a subtle structural shift. It does not take much — fractions of a pixel of perceived position difference are enough to shift the face from rendered to photographed.
For more on the structural reasons this happens across AI images broadly, see make AI photos look real in 2026.
6.Hair: From Illustrated to Photographed
AI hair fails at two scales. At the macro scale, volume is too uniform — every strand appears equidistant, as if sculpted. At the micro scale, there are no flyaways, no individual strands catching the light at a different angle from their neighbours, no clumping along the hairline.
Real photographed hair has all three of these things. When you look at a high-resolution portrait, you can see individual strands at the hairline, slight variations in sheen between different parts of the crown, and the way a few stray hairs separate from the rest and catch the background light.
Prompting for hair realism:
- 'Individual hair strands visible at the hairline'
- 'Natural flyaways, slight windswept texture'
- 'Realistic hair sheen with strand-to-strand variation'
- 'Hair slightly lifted by light breeze'
Lighting is the other lever. Flat frontal lighting makes all strands read as one surface. A light source at 45 degrees from behind — a hair light or rim light — forces the model to compute shadows between individual strands, which is the physical cue that makes hair look three-dimensional and photographed rather than painted.
7.Putting It Together: The Portrait Prompting Stack
For a realistic AI portrait in 2026, build your prompt in this sequence:
- Light source first: 'Soft window light from camera left, warm afternoon, slight shadow fill on right cheek'
- Camera and lens: 'Shot on Sony A7IV, 85mm f/1.4 portrait lens, shallow depth of field, eyes in tack-sharp focus, soft background bokeh'
- Skin texture: 'Realistic skin pores, natural subsurface scattering, slight skin oiliness, faint freckles'
- Structural realism: 'Natural facial asymmetry, slightly uneven features, natural expressions'
- Hair detail: 'Individual hair strands, natural flyaways at hairline, realistic hair sheen'
- Expression: 'Candid mid-expression, natural relaxed smile, looking slightly off-camera'
The 85mm focal length specification is worth noting: it is the classic portrait focal length because it produces gentle feature compression and natural perspective. Wide-angle focal lengths (below 35mm) distort facial proportions in ways that exaggerate the uncanny quality of an AI face. Always specify a portrait-appropriate focal length.
After generation, run the face region through Imagera's portrait upscaler to recover iris texture, pore detail, and hair strand separation. Browse plan options at /pricing — portrait quality enhancement is available from the Starter plan.
