AI Photo Realism Looks Plastic: Fix the Fake Skin Look

Q: My faces still look plastic on Flux 2 even with good prompts. What now?

Lower the guidance (CFG) to around `3.5`. High guidance pushes Flux toward over-smoothed, waxy skin. Pair the lower guidance with `natural skin texture, avoid over-smoothing` and a real lens reference.

Photorealistic prompts come out looking plastic and fake. Fix it with concrete lighting, skin-texture, lens, and negative-prompt cues — verified June 2026.

Published: May 17, 2026 Updated: Jun 17, 2026 Author: AI Productivity Guide Team 🌐 查看中文版本

You wrote “photorealistic portrait of a woman,” picked a strong realism model, and the output looks like a video-game render: skin is too smooth, hair has no flyaways, lighting is too even, eyes have a glassy doll-like quality. This is the “AI plastic” look. It is not a model limitation per se; it is what models default to when the prompt gives them no anchors for real-world imperfection.

Fastest fix (works most of the time): delete the word photorealistic and every quality token (8k, hyperdetailed, masterpiece), then add one concrete lighting setup plus one real lens. For example, replace photorealistic portrait, 8k, hyperdetailed with soft window light from camera left, shot on 85mm f/1.4, visible skin pores, fine flyaway hairs. That one swap closes most of the gap. The steps below add the rest.

Which bucket are you in?

Symptom you see	Most likely cause	Jump to
Flat, evenly-lit, no shadow direction	No lighting specification	Step 1
Skin is a smooth gradient, no pores	Missing texture descriptors / negative prompt	Step 2 + Step 4
Generic deep-focus “AI portrait” feel	No camera / lens language	Step 3
Glossy, “asset-store,” over-polished	Stacked quality tokens (`8k`, `masterpiece`)	Step 4
Uncanny perfectly-symmetric face	Symmetric default	Step 5
Stylized/illustration look no matter what	Wrong base model	Step 6
Looks “too clean,” not like a real file	No post-process (grain, color)	Step 7

Common causes

Ordered by what most often produces the plastic look.

1. No lighting specification

photorealistic alone is a meaningless instruction. Every real photo has a specific lighting setup: window light, golden hour, overcast, studio softbox, hard noon sun. Without that anchor, the model averages to a flat, evenly-lit composite that immediately reads as fake.

How to spot it: Search your prompt for any phrase describing light source, direction, color temperature, or time of day. If none, this is your problem.

2. “Photoreal” / “hyper-realistic” alone

These modifiers are so overused that models have learned them as code for “the polished, plastic, asset-store look.” Saying photoreal can paradoxically produce less realistic output than saying nothing.

3. Missing texture / material descriptors

Real skin has pores, oil sheen, freckles, fine hairs. Real fabric has weave, wrinkles, fiber shadow. Real wood has grain and minor surface variation. Models default to a sanitized version of all of these unless prompted otherwise.

How to spot it: Zoom into a 100x100 patch of skin in your output. Is it a uniform smooth gradient? Real skin has visible high-frequency texture at this zoom level.

4. No camera / lens language

shot on iPhone and shot on Canon R5 with 85mm f/1.4 produce dramatically different outputs. Without lens language, the model defaults to a generic, deep-depth-of-field, vaguely-flattering portrait look that screams “AI render.”

5. Over-symmetric, perfect-feature subject

Real faces are slightly asymmetric. AI defaults to symmetric. Prompting perfectly symmetrical, or just leaving the default, makes the output uncanny.

6. Wrong base model for the task

Some models are trained heavily on illustration and stylized data (Niji, anime-focused SDXL checkpoints). Asking them for realism produces stylized output regardless of prompt. Use a realism-focused base. As of June 2026 the strongest options are Flux 2 (Black Forest Labs), Imagen 4 (Google), and on the open-source side Juggernaut XL and RealVisXL.

7. Too many style modifiers

photorealistic, cinematic, dramatic, hyperdetailed, 8k, masterpiece: every one of these tokens is associated with the AI-render look in the training data. The more you stack them, the more “AI” the output looks.

Before you change anything

Save the current prompt, model, version, seed, and the plastic-looking output.
Find one reference photo (real, not AI) of roughly what you want. Note its lighting, framing, and texture cues.
Decide whether the use case can tolerate any AI tell, or whether it must pass close inspection.
Confirm you are using a realism-focused base model, not a stylized one.
Back up the current prompt template before changing it.

Information to collect

Full prompt, negative prompt, model, version, seed, and (for SD-family) CFG/guidance value.
A 100x100 crop of the output’s skin / surface area.
The reference real photo you are trying to match.
The intended use case (social, print, hero on a landing page).

Shortest path to fix

Step 1: Specify lighting in concrete photographic terms

Replace photorealistic with a lighting setup:

soft window light from camera left, slight fall-off, warm golden hour,
shot in the late afternoon

Other strong lighting anchors:

overcast diffuse light, even shadows
studio softbox key light, rim light from behind
single hard noon sun, deep shadows
candle light, low key, dim ambient

This single change often produces half of the realism gain.

Step 2: Add specific texture descriptors

For people:

visible skin texture, fine pores, subtle freckles, individual flyaway hairs,
natural skin oil reflection on the nose and forehead

For products / materials:

matte brushed aluminum with fine machining marks, slight fingerprint on surface

worn leather with visible grain and slight creasing

Step 3: Use camera + lens language

shot on Canon R5 with 85mm f/1.4 lens, shallow depth of field,
slight grain, color profile Kodak Portra 400

Substituting a real lens for shallow depth of field produces noticeably different results because the model has learned per-lens characteristics from photographer training data. Two camera references that reliably anchor realism: Hasselblad X2D for an editorial/studio look, and iPhone rear camera for a candid, snapshot feel.

Step 4: Drop AI-tell modifiers, and (where supported) add a skin negative prompt

Delete from your prompt:

8k
hyperdetailed
ultra realistic
masterpiece
award winning
trending on artstation

These are correlated with the polished AI-render look. Counterintuitively, your output gets more realistic when you remove them.

For models that support a negative prompt (Stable Diffusion family: SDXL, Juggernaut XL, RealVisXL), push back on the plastic tendency directly:

(airbrushed:1.3), (smooth skin:1.2), plastic, wax, doll, 3d render,
porcelain skin, beauty filter, flat lighting

Important: Flux 2 and Imagen 4 do not use a negative prompt at all. With those, describe what you want instead of what you want to avoid. Write natural skin texture, avoid over-smoothing in the positive prompt rather than relying on a negative.

Step 5: Add slight imperfection cues

slight asymmetric features, one eyebrow slightly higher, subtle blemish
on the left cheek, natural human imperfection

This forces the model away from its symmetric default.

Step 6: Switch to a realism-focused model (verified June 2026)

If steps 1-5 still produce a plastic look, the model is the bottleneck. Current strongest options:

Flux 2 (Black Forest Labs) — leads photorealism in 2026 for skin texture and lighting accuracy. No negative prompt; if faces still look plastic at high guidance, drop CFG/guidance to about 3.5 and add natural skin texture, avoid over-smoothing.
Imagen 4 (Google) — very strong on faces and complex scenes; available via Google AI Studio and Vertex AI. Also no negative prompt.
Juggernaut XL (v10) or RealVisXL (V5.0) — best open-source realism in the SD family; use the skin negative prompt from Step 4.
Midjourney V8.1 (default on midjourney.com since June 10, 2026; V7 still selectable) — add --raw to strip the default aesthetic and let your camera/lighting cues carry the look. For portraits keep stylize low, around --s 50 (the default is --s 100, range 0 to 1000).

Step 7: Post-process to add film grain and color shift

A subtle film grain (5-10% intensity) plus a gentle color shift in DaVinci Resolve or Lightroom adds the final layer of believability. Real digital photos are processed; pure AI output is not.

How to confirm the fix

Zoom to 100% on skin or surface — you should see visible high-frequency texture, not a smooth gradient.
Show the output to someone who does not know it is AI. If they assume it is a photo, you have arrived.
Check that the lighting direction and color in the output match what your prompt specified (shadows fall the right way).
Side-by-side with your real reference photo, the gap should be narrowing, not still glaringly different.

If it still fails

Strip the prompt to its minimum (subject + one lighting cue + one lens cue), regenerate. Add back one descriptor per cycle and watch which one re-introduces the plastic look.
Use image-to-image starting from a real photo (or another realistic AI image), with denoise 0.5-0.6, instead of pure text-to-image.
Run a face / skin restoration pass: ADetailer with a skin-detail LoRA at weight 0.4-0.6 (full strength can make skin look like sandpaper), or a Photoshop high-pass on skin.
Generate at higher resolution (Flux at 1536x1536, SDXL with Hires Fix at 1.5x). More pixels means more room for texture detail.
Package the prompt, model, version, output crops, and reference photo before asking for community help.

FAQ

Why does adding “photorealistic” make my image look worse? Because the token is so overused in training data that models associate it with the polished, over-smoothed render look. Replace it with concrete lighting and a real lens; those carry far more realism signal.

Should I use a negative prompt to kill plastic skin? Only on models that support one. For the SD family (SDXL, Juggernaut XL, RealVisXL) a weighted negative like (airbrushed:1.3), (smooth skin:1.2), plastic helps. Flux 2 and Imagen 4 ignore negative prompts entirely; with those, describe natural skin texture positively instead.

My faces still look plastic on Flux 2 even with good prompts. What now? Lower the guidance (CFG) to around 3.5. High guidance pushes Flux toward over-smoothed, waxy skin. Pair the lower guidance with natural skin texture, avoid over-smoothing and a real lens reference.

What is the single best model for realistic portraits right now (June 2026)? Flux 2 for skin texture and lighting realism, Imagen 4 for complex multi-subject scenes, and Midjourney V8.1 with --raw for an aesthetic-but-believable look. On open-source, Juggernaut XL v10 or RealVisXL V5.0.

The whole face is right but it reads as “too clean.” What is missing? The post-process layer. Real digital photos have grain, slight color grading, and minor optical imperfection. Add 5-10% film grain and a small color shift in Lightroom or DaVinci Resolve.

Prevention

Build a “realism kit” of phrases that work for your favorite model: lighting setups, lens, film stocks, and reuse them.
Always lead a realism prompt with concrete lighting and lens descriptors, not photoreal.
Keep a folder of reference real photos by use case and reference them when writing prompts.
Standardize on a realism-focused base model for any portrait or product work.
Apply post-process film grain to any AI output destined for print or premium placement.

Tags: #Prompt #Debug #Troubleshooting #Image generation #Realistic