How to Generate Realistic AI Images: Fix Plastic Skin & Fake Portraits
You generate an AI portrait and the result looks like a shiny mannequin. The skin is flawless, the lighting is flat, and the eyes have no life. You try again with a slightly different description. Same result.
This is not a Nano Banana limitation — it is a prompting problem. The model defaults to the statistical average of its training data, which skews heavily toward over-retouched, airbrushed stock photography. Every undefined detail in your prompt gets filled with the most "polished" version the model has learned.
This guide gives you the exact tactics to override that default and force Nano Banana to output raw, candid, skin-texture-accurate portraits.
Quick answer: To fix plastic-looking AI skin, add
visible skin pores+ a real camera reference (Hasselblad X2D for editorial, iPhone rear camera for candid) + directional lighting + film grain tokens + negative constraints (no smooth skin filter, no beauty retouch, no symmetrical face). Each of these targets a different layer of the model's beautification default.
Why AI Defaults to "Plastic"
The core issue is what researchers call the uncanny valley of skin rendering. AI models trained on commercial photography learn that "beautiful portrait" equals smooth skin, even lighting, and symmetrical features — because that is what millions of edited stock images look like.
When you write "a beautiful woman looking at the camera" without additional constraints, you are not prompting for a person. You are prompting for the model's average of every retouched headshot it has ever seen.
The fix is not to add more adjectives. The fix is to introduce the specific photographic imperfections that signal "real camera, real person, no post-processing."
6 Tactics That Eliminate Plastic Skin
1. Command Skin Texture Explicitly
Never rely on "realistic" alone. It means nothing to the model after millions of uses. You need specific texture terms:
visible skin poresnatural pore densitypeach fuzz on cheek and jawslight hyperpigmentationdry skin texture at the sides of the noseunretouched, no frequency separation
Each of these tokens triggers a different micro-texture calculation. Together, they force the model to render skin at a cellular level instead of smoothing it into a CGI surface.
2. Add Catchlights to the Eyes
Nothing kills portrait realism faster than flat, expressionless eyes. In real photography, the light source always reflects as a small specular highlight inside the iris — called a catchlight. Without it, the subject looks dead.
Add these to every portrait prompt:
cinematic catchlight in eyesnatural specular reflection in irispin-sharp eyes with visible catchlight
The direction matters too: catchlight from upper-left matches a window-lit portrait. Dual catchlights, softbox setup reads as studio. Matching the catchlight to your stated light source makes the whole image more coherent.
3. Emulate Real Camera Hardware
The lens and camera body you specify directly changes how Nano Banana renders micro-detail. This is not flavor text — it is a semantic trigger.
| Camera reference | Effect on output |
|---|---|
| iPhone 15 Pro, rear camera, candid | Slight digital noise, compressed JPEG feel, mobile authenticity |
| 35mm f/1.8, slight vignette | Street-photography grain, warm color science |
| 85mm f/1.4, shallow depth of field | Portrait compression, subject-background separation |
| Hasselblad X2D, 135mm, medium format | Dense skin texture, extreme dynamic range, studio authority |
The medium-format reference (Hasselblad, Fujifilm GFX) consistently produces the most textured, pore-accurate skin rendering in Nano Banana — because those cameras are associated with unretouched editorial photography in the training data.
4. Use Analog Tokens to Block Digital Post-Processing
AI models have an internal tendency to "beautify" — to add a digital sheen or glow that reads as post-processed. You can debuff this behavior with analog-era tokens that imply the absence of digital intervention:
slight film grain, ISO 800subtle JPEG compression artifactshot on Kodak Portra 400unedited RAW export, no color gradeslight chromatic aberration at edges
The ISO grain token is particularly effective — it introduces luminance noise that breaks up the perfectly smooth gradients the model naturally wants to produce.
5. Define Directional Lighting — Never "Good Lighting"
Flat, undefined lighting is a major source of the plastic look. When there are no hard shadows or directional cues, skin loses all three-dimensional texture.
Replace vague lighting with specific setups:
soft window light from the left, natural shadow falling across right side of faceovercast outdoor light, even diffusion, slight skin texture revealedraking sidelight, 45-degree angle, strong shadow definition on skin texturemorning backlight, rim light on hair, face in open shade
Raking light (light coming from a low, side angle) is the most aggressive texture-revealing setup. It is standard in commercial skin photography precisely because it catches every pore and surface irregularity.
6. Negative Constraints — The Layer Most People Skip
Positive prompts describe what you want. Negative constraints tell the model what it is explicitly forbidden to do. Nano Banana responds strongly to both.
For every portrait targeting realism, include:
no smooth skin filter, no beauty retouch, no plastic skin texture,
no symmetrical face, no airbrushed look, no HDR glow,
no digital sheen, no over-saturated colors, no AI aesthetic
The no symmetrical face constraint is particularly important — human faces have natural asymmetry, and the model's tendency to produce perfect bilateral symmetry is one of the most obvious tells of an AI-generated image.
A Complete Prompt: Before and After
Before (typical user prompt):
A 30-year-old woman with brown hair, looking at the camera, natural light, realistic.
After (full texture-locked prompt):
30-year-old woman, slight dark circles, asymmetrical face, loose brown hair — looking directly at camera, soft morning light from the left, natural shadow on right side — 85mm f/1.4, shallow depth of field — visible pore density, peach fuzz on jaw, slight dry texture at nose sides, pin-sharp eyes with cinematic catchlight from upper-left — shot on Hasselblad X2D, slight film grain ISO 400, unretouched RAW — no smooth skin filter, no beauty retouch, no symmetrical face, no digital sheen.
The difference in output is not incremental — it is categorical. The second prompt leaves the model zero room to default to its polished average.
These tactics apply within a larger prompting system. The complete 8-part prompt formula shows how all layers — subject, outfit, environment, camera, lighting, texture, and negatives — interact to produce consistently realistic output. If you are working from a reference photo, the image-to-prompt reverse engineering guide explains how to extract these parameters automatically from any image.
Automating This With CookedBanana
Writing a full texture-locked portrait prompt for every generation takes 5–10 minutes of deliberate work. CookedBanana compresses this to under 10 seconds.
Upload your reference photo — the candid look you want to replicate — and the engine automatically extracts lighting conditions, skin texture cues, camera simulation, and negative constraints, formatting them into a structured prompt ready for Nano Banana. The 8-macro-reference system ensures every layer (subject, lighting, hardware, texture, negatives) is covered without missing a single critical term.
Start your free trial — 3 generations, no credit card required.
Frequently Asked Questions
Why does Nano Banana always make skin look too smooth?
Because it is defaulting to the statistical average of its training data, which skews toward heavily retouched commercial photography. Without explicit texture terms like visible pores, peach fuzz, or natural skin grain, the model interprets "realistic" as "idealized." The fix is to specify the imperfections directly — the model knows what they look like, it just needs permission to render them.
What is the single most effective prompt keyword to fix plastic skin?
Visible skin pores consistently produces the most immediate improvement across portrait types. It forces the model to calculate micro-surface detail rather than smoothing it. For maximum effect, combine it with a strong directional lighting instruction and a medium-format camera reference.
Does the camera reference in the prompt actually matter?
Yes — significantly. The camera and lens tokens are semantic triggers that associate with specific photographic traditions in the training data. A Hasselblad X2D reference associates with unretouched medium-format editorial photography. An iPhone 15 Pro, candid reference associates with unedited mobile photography. Both produce more authentic skin texture than generic "photorealistic" tags.
What are catchlights and why do they matter for AI portraits?
A catchlight is the small specular reflection of a light source visible inside the subject's iris. In real photography it is always present — and its absence is one of the most immediate signals of an AI-generated or over-retouched image. Adding cinematic catchlight in eyes or natural specular reflection in iris to your prompt forces Nano Banana to include it, giving the subject visible life and presence.
Can I use these tactics for full-body portraits, not just close-ups?
Yes, with adjustments. For full-body portraits, shift the texture focus from the face to clothing and environment: fabric texture on garment, visible thread detail, natural wrinkles on linen, dirt on shoe sole. The principle is identical — introduce the specific physical imperfections that signal an unedited, real-world capture rather than a digitally idealized render.
How do I fix AI eyes that look glassy or lifeless?
Glassy eyes are caused by missing specular detail and incorrect iris rendering. Add cinematic catchlight in eyes, natural specular reflection in iris, pin-sharp eyes, and visible iris texture. Also check your lighting definition — the catchlight direction should match the stated light source (e.g., catchlight from upper-left pairs with soft window light from the left).
What is the difference between film grain and digital noise in AI prompts?
Film grain (slight film grain ISO 800, shot on Kodak Portra 400) associates with analog photography and produces organic, warm texture that reads as pre-digital. Digital noise (ISO 1600 digital noise, mobile sensor noise) reads as smartphone candid and works better for UGC or lifestyle content. Both override the model's tendency to produce overly clean gradients — but they produce distinct aesthetic results.
Topics