ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›WHY DO AI IMAGES LOOK PLASTIC? HOW TO FIX
ARTICLES · HOW-TOBEGINNER

Why Do AI Images Look Plastic? How to Fix Glossy, Waxy Skin

Why AI skin looks plastic, waxy or glossy, and how to fix it: prompt rewrites and settings for ChatGPT, Gemini, Midjourney, FLUX.2 and Stable Diffusion.

READ16 min
WORDS3,143
SECTIONS12
TYPEHOW-TO
CHECKED13 OCT 26

AI images look plastic when the model gives skin one smooth, even, slightly shiny surface with no pores, fine lines or small marks. The fix is to ask for real texture by name, light from one side, camera and film words, lower style or guidance settings, and a gentle upscale.

It happens partly because some image models were tuned toward polished pictures people like, and partly because flat light, “perfect” prompt words, high style settings, smoothing upscalers and repeated edits all push the same way.

TL;DR — THE SHORT VERSION
  • Describe the skin you want: visible pores, faint fine lines, freckles, a slightly uneven tone. “No plastic” on its own does little.
  • Light from one side, so it skims across the skin and shows its texture. Flat, even light hides it.
  • Use camera and film words instead of “8K, masterpiece, flawless”.
  • Turn style down: lower stylize or --raw in Midjourney, a lower guidance setting in Stable Diffusion.
  • Upscale gently (Midjourney’s Subtle upscaler, a low strength or creativity setting), and edit from the original, not from an edit of an edit.

What does “plastic” look like in an AI image?

The plain point: the skin looks like a doll’s or a wax figure’s: smooth, evenly lit and faintly glossy.

  • No pores or fine lines, even on a close-up of an older face.
  • One even glow across the forehead, cheeks and nose, like a sheen of lotion.
  • No small marks: no freckles, moles, redness, stubble or fine hair.
  • Too-perfect symmetry, with a face that looks retouched for an advert.

Other things can look “off” too, such as light from two directions or a stock-photo pose. Those are covered on why AI output looks fake. This page is about the skin.

Why do AI images look plastic?

The plain point: six things push skin toward smooth and shiny. Some are documented; others are this site’s reading from use, and are labelled as such.

  1. The model learned to please. Some image models were trained on pictures picked for how much people would like them. LAION built its LAION-Aesthetics image sets with a model that predicts how people answered “How much do you like this image on a scale from 1 to 10?”, and says Stable Diffusion’s first version was trained on one of those sets.1 Krea, a company that trains image models, found that this LAION filter is biased toward soft textures, blurry backgrounds and bright images, and that picking training pictures this way “adds implicit biases to the model’s priors”: what the model leans toward by default. It counts “waxy skin textures” as part of the “AI look”.2 Other makers publish less about how they pick training images.Reasoning, 13 Oct 2026 — LAION and Krea document this for LAION-Aesthetics and the models trained with it; this site found no comparable statement from other image-model makers.
  2. Flat, front-on light. Light from the front fills in every tiny shadow on the skin, and those shadows are what show texture. Light from the side does the opposite.Reasoning, 13 Oct 2026 — this site’s reading from how light and texture behave in photography, not a vendor claim.
  3. “Perfect” prompt words. Words like flawless, beautiful, glowing, perfect skin and masterpiece describe a polished advert look, so that is what comes back.Reasoning, 13 Oct 2026 — this site’s reading from use, not a measured test. Piling them up does not help either: a Civitai community guide warns that “Repeating the same keywords does not necessarily improve results and may lead to distortions.”3
  4. High style and guidance settings. Midjourney says a higher stylize value gives it “more freedom to interpret your idea” and that the image “might look more artistic”.4 For Stable Diffusion, Hugging Face warns that if the guidance value “is too high, you may observe some artifacts in the generated image”.5 In this site’s use, an over-styled or over-guided face is often the waxiest one.
  5. Upscalers and face restorers. Some upscalers redraw the picture as they enlarge it. Midjourney says its Creative upscaler “will change your image a bit as it adds new details”.6 The face restorer CodeFormer describes a trade-off between a “higher-quality” result and a “higher-fidelity” one, meaning truer to the photo you put in.7 When a redraw or a restorer leans toward a clean, “nice” face, real texture is often what gets lost (this site’s reading).
  6. Editing an edit, again and again. OpenAI warns that “Repeated edits can still change details you intended to preserve.”8 In this site’s use, skin texture is one of the first details to fade over many rounds.
◈ IN PLAIN TERMS

Think of a photo that went through a beauty app, then a second app, then a third. Each one tidies a little more until the face looks like a shop dummy. An AI image can get the same treatment from its training, your words, its settings and its upscaler, all before you see it.

How do you fix plastic skin in any tool?

The plain point: say what real skin looks like, light it from one side, and talk like a photographer.

  • Name the texture. Visible pores, faint fine lines, freckles, a few small marks, a slightly uneven skin tone, fine hair on the arms. OpenAI’s own example portrait asks for “weathered skin with visible wrinkles, pores, and sun texture” and “real skin texture”.8 More words are on micro detail prompts.
  • Light that rakes across the skin. Window light from one side, low sun from the side, or one lamp off to the left. “Rakes” means the light skims across the surface at a low angle, so every tiny bump casts a tiny shadow. More lighting words: colour, light and mood.
  • Camera and film words. Black Forest Labs says: “For photorealism, specify camera models, lenses, and film stocks.”9 OpenAI adds a caution: treat camera details “as cues for appearance, not a guarantee of exact physical simulation”.8 See camera grammar.
  • Drop the “perfect” words. Flawless, glowing, porcelain, 8K, masterpiece. Realism words explains why “hyperrealistic” points away from a photograph.
  • Ask for no retouching. “Unretouched”, “no beauty filter”, “no heavy retouching”. On tools with a negative field, put the unwanted look there instead (see the tool sections below).
TAKEAWAY

“No plastic skin” tells the model what to avoid. “Visible pores, window light from the left, 35mm film” tells it what to make. The second works on every tool.

Before and after: prompt fixes to copy

The plain point: each “before” is a prompt that tends to come out plastic; each box below it is the rewrite. Swap the bracketed parts for your own subject.

1. Before: beautiful woman, perfect skin, flawless, ultra realistic, 8K, masterpiece

USE WHEN: A CLOSE-UP PORTRAIT COMES OUT WAXY
Close-up portrait photo of a [woman in her thirties] sitting by a window, natural skin texture with visible pores, faint freckles across her nose, a few fine lines at the eyes, slightly uneven skin tone, no makeup, soft daylight from the left, shot on a 50mm lens at eye level, unretouched

2. Before: realistic old man portrait, highly detailed

USE WHEN: AN OLDER FACE LOOKS AIRBRUSHED
Documentary photo of a [man in his seventies] on a park bench, weathered skin with deep wrinkles, sun spots and visible pores, grey stubble, low afternoon sun raking across his face from the right, 35mm film photograph with subtle grain, natural colours

3. Before: portrait, studio lighting, beautiful glow

USE WHEN: THE LIGHT IS FLAT AND EVEN
Portrait lit by a single window to camera left, the light skimming across the cheek so the skin texture shows, the far side of the face falling into soft shadow, no fill light, no glow

4. Before: glowing skin, dewy, radiant

USE WHEN: THE SKIN SHINES LIKE PLASTIC
matte skin with a faint natural shine only on the forehead and the tip of the nose, visible pores on the cheeks, no highlighter, no beauty filter

5. Before: hand holding a coffee cup, perfect hands

USE WHEN: HANDS AND ARMS LOOK LIKE A MANNEQUIN’S
Close-up of an adult hand holding a ceramic coffee cup, visible knuckle creases, fine hairs on the back of the hand, short unpolished nails, slightly dry skin, morning light from the side, shallow depth of field

6. Before: professional photo of friends at a cafe, 8K, sharp

USE WHEN: YOU WANT AN ORDINARY, REAL-LOOKING SNAPSHOT
Candid phone photo of two friends laughing at a cafe table, taken at arm’s length, slightly off-centre, mixed indoor light, real skin texture, a little motion blur on one hand, ordinary unedited colours

7. Before: make the skin look more realistic (as an edit)

USE WHEN: EDITING A FINISHED IMAGE IN CHATGPT OR GEMINI
Edit this image. Change only the skin: bring back natural texture with visible pores, faint fine lines and a slightly uneven tone, as in an unretouched photograph. Keep the face shape, features, expression, hair, lighting and background exactly the same. Do not smooth, sharpen or beautify anything.

8. Before: upscale this, make it high quality

USE WHEN: UPSCALING A PORTRAIT WITH A PROMPT
Upscale this image. Keep the skin texture exactly as it is, including pores, fine lines, freckles and small marks. Do not smooth, retouch or beautify the skin, do not add a glow, and do not change the face, expression or lighting.

ChatGPT Images: describe texture and state the exclusion

The plain point: exclusions go inside the prompt, not in a separate box, so the fix lives in your words.

OpenAI’s image prompting guide covers its GPT Image models, and OpenAI says the newest API model “brings the same improvements” as ChatGPT Images 2.5.10 The guide’s advice for realism: ask for “photorealistic” or “real photograph” and “describe framing and texture”.8 Exclusions go inside the prompt: OpenAI’s guide says “State exclusions such as unwanted text, logos, or watermarks.”8 Its own realistic portrait example ends with “No glamorization, no heavy retouching.”, a line you can borrow.8

USE WHEN: CHATGPT IMAGES — ADD AT THE END
Real photograph, natural skin texture with visible pores and fine lines. No glamorization, no heavy retouching, no beauty filter.

When editing, change one thing per request and restate what must stay the same, as OpenAI advises: “restate critical constraints if the result drifts”.8 If the skin has gone waxy after many rounds, start again from the original. Formula and limits: ChatGPT Images cheat sheet.

Gemini (Nano Banana): describe the scene, light and lens

The plain point: Google asks for a described scene, not a keyword list, and has no negative box.

Google’s template for photorealistic scenes includes the light and the lens: “A photorealistic [type of shot] of a [subject description] in a [setting description]. [Description of the light]. Shot from a [camera angle] with a [lens type].”11 Fill the light slot with side light and the subject slot with texture. Instead of a ban, Google asks for a “semantic negative prompt”: describe what you want present.11 Google also suggests small follow-ups such as “Keep everything the same, but change the character’s expression to be more serious.”11 The same pattern works for skin:

USE WHEN: GEMINI — FOLLOW-UP MESSAGE
Keep everything the same, but give the skin natural texture: visible pores, faint fine lines and a slightly uneven tone, with the light coming from the left side.

More rules, quoted: Nano Banana cheat sheet.

Midjourney: --raw, lower stylize and single-word --no

The plain point: Midjourney has three settings that help: Raw, a lower stylize value and the --no list.

  • --raw. Midjourney says Raw turns off its own creative touch: “With simple prompts, you’ll often get more realistic, photo-like images.”12 Older prompts often write it --style raw; Midjourney’s current page gives --raw (this site’s note).
  • Lower --stylize. The default is 100 and the range is 0 to 1000. Midjourney says a low value means “Midjourney sticks to the facts without much extra flair.”4
  • --no with single words. Midjourney says it reads every word in --no on its own, so “--no modern clothing” is read as “no modern” and “no clothing”.13 By the same logic, “--no plastic skin” risks reading as “no skin”, so use words like airbrushed, retouched and glossy (this site’s reading).
USE WHEN: MIDJOURNEY PORTRAIT
portrait of a [fisherman in his sixties] on a harbour wall, weathered skin, visible pores, grey stubble, overcast light from one side, 35mm film photo --raw --s 50 --no airbrushed, retouched, glossy

For upscaling, pick Subtle rather than Creative (see below). All parameters: Midjourney cheat sheet.

FLUX.2: no negatives, so describe the film and the light

The plain point: FLUX.2 ignores “no” lists, so every fix must be written as something present.

Black Forest Labs says “FLUX.2 does not support negative prompts” and gives the example: “Instead of ‘no blur,’ say ‘sharp focus throughout.’”9 Its style guide lists film looks you can borrow, such as “shot on Kodak Portra 400, natural grain, organic colors”.9 On FLUX.2 [flex], BFL lists a guidance setting for “Prompt adherence”, default 4.5, range 1.5 to 10.9 BFL does not say a lower value fixes skin; in this site’s reading it is worth one test run a step lower if faces look over-processed.

USE WHEN: FLUX.2 PORTRAIT
Portrait of a [ceramicist in her fifties] at a pottery wheel, natural skin texture with visible pores and fine lines, clay dust on her forearms, window light from the left, shot on Kodak Portra 400, natural grain, organic colors

More: FLUX.2 cheat sheet.

Stable Diffusion and SDXL: negative list and lower guidance

The plain point: Stable Diffusion has a real negative prompt box and a guidance number you set yourself.

Hugging Face describes the negative prompt as “The prompt or prompts to guide what to not include in image generation.”14 Paste a plastic-skin list there:

NEGATIVE — PLASTIC OR WAXY SKIN (STABLE DIFFUSION, SDXL)
plastic skin, waxy skin, airbrushed skin, beauty filter, over-smoothed skin, porcelain skin, doll-like face, glossy skin, heavy retouching, CGI skin

The guidance setting (see CFG scale) controls how hard the model follows your prompt. For SDXL, the Hugging Face docs give a default of 5.0 and say a higher value works “usually at the expense of lower image quality”.15 If skin looks waxy and your guidance is well above the default, bring it back toward the default and compare (this site’s advice). The negative box only works while guidance is on: Hugging Face says it is “Ignored when not using guidance”.14 See the plastic skin negatives and the negative prompt builder.

How to upscale without making skin plastic

The plain point: choose the setting that changes the least, then check the skin at full size.

  • Midjourney: use Subtle. Midjourney says it “keeps changes to a minimum, preserving the look of your initial image as closely as possible”, while Creative “adds new details”.6
  • Stable Diffusion image-to-image: keep strength low. Hugging Face says “a lower strength value means the generated image is more similar to the initial image”.16 Some apps call a similar setting denoise (this site’s note).
  • Other upscalers with a “creativity” or “resemblance” slider: start with low creativity and high resemblance, and turn off face enhancement for portraits (this site’s advice; check each tool’s own help page for what its slider does).
  • Prompt-based upscaling in ChatGPT or Gemini: use fix 8 above, which tells the model to keep the texture.

Upscaling makes a picture bigger; it does not make plastic skin real. Fix the skin first, then upscale. See upscaling meaning and upscale prompts.

Questions people ask

Why does AI skin look so smooth?

Partly because some models were trained on images people rated as likeable, which LAION documents for Stable Diffusion’s first version.1 Krea, which trains image models, found that filter leans toward “overly soft textures”.2 Flat light, “perfect” words and smoothing upscalers add to it, in this site’s reading.

Does “8K” or “ultra realistic” fix plastic skin?

Not in this site’s use: they describe a polished look, not texture. Vendors point to description instead; OpenAI asks you to “describe framing and texture”.8 See realism words.

Can a negative prompt fix plastic skin?

Only on tools that read one, such as Stable Diffusion and Midjourney’s --no. FLUX.2 does not support negative prompts, so there you describe the texture instead.9

Is “--style raw” still a Midjourney setting?

Midjourney’s current docs list Raw as --raw, compatible with versions 5.1 and later.12 You can also switch it on for all prompts in the settings panel.

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. Which change does the page say works on every AI image tool to fix plastic skin?

  • Adding 8K, masterpiece and flawless to the prompt
  • Describing real skin texture and lighting it from one side
  • Running the image through a creative upscaler twice
Show the answer

Answer: Describing real skin texture and lighting it from one side. The page says naming the texture you want and using light from one side works everywhere, while perfect words and creative upscaling push toward smooth skin.

2. Why does the page suggest single words like airbrushed and glossy after Midjourney’s --no?

  • Midjourney reads each word in --no on its own, so plastic skin could read as no skin
  • Midjourney only accepts one word per prompt
  • Single words make the image generate faster
Show the answer

Answer: Midjourney reads each word in --no on its own, so plastic skin could read as no skin. Midjourney says it reads every word in --no independently, so a two-word phrase can be split in a way you did not mean.

3. When upscaling a portrait, which setting does the page recommend to keep skin texture?

  • The highest creativity setting, so it adds new detail
  • Face enhancement turned on for every portrait
  • The option that changes the least, such as Midjourney’s Subtle upscaler or a low strength
Show the answer

Answer: The option that changes the least, such as Midjourney’s Subtle upscaler or a low strength. The page says to pick the setting that changes the least, because creative upscalers and face restorers can redraw and smooth the skin.

SOURCES
  1. 1LAION, LAION-Aesthetics, 16 Aug 2022, read at source 13 Oct 2026“we trained several lightweight models that predict the rating people gave when they were asked ‘How much do you like this image on a scale from 1 to 10?’” and “They used the 5+ subset to train Stable Diffusion V1 model.” ↑a“They used the 5+ subset to train Stable Diffusion V1 model.” ↑b
  2. 2Krea, Releasing Open Weights for FLUX.1 Krea, 31 Jul 2025, read at source 13 Oct 2026“we find that LAION Aesthetics — a model commonly used to obtain high quality training images — to be highly biased towards depicting women, blurry backgrounds, overly soft textures, and bright images.” and “relying on these models to obtain high quality training images adds implicit biases to the model’s priors.” and the “AI look” is “overly-blurry backgrounds, waxy skin textures, boring composition, and more.” ↑a“we find that LAION Aesthetics — a model commonly used to obtain high quality training images — to be highly biased towards depicting women, blurry backgrounds, overly soft textures, and bright images.” ↑b
  3. 3Civitai Education, Civitai’s Prompt-Crafting Guide: Part 1 – Basics, last updated 13 Mar 2025, read at source 12 Oct 2026“Repeating the same keywords does not necessarily improve results and may lead to distortions.” ↑
  4. 4Midjourney, Stylize, read at source 13 Oct 2026“With a higher stylize setting, it’s like giving Midjourney more freedom to interpret your idea. The image might look more artistic and visually interesting, but it could stray from the exact details of your prompt.” ↑a“With a low stylize setting, it’s like asking for an image that follows your prompt very closely—Midjourney sticks to the facts without much extra flair.” and “The default value for stylize is 100, and you can adjust it anywhere between 0 and 1000 with the latest Midjourney versions.” ↑b
  5. 5Hugging Face, Diffusers: Text-to-image, read at source 13 Oct 2026“Higher guidance_scale values push the model to follow the prompt more closely, and if this value is too high, you may observe some artifacts in the generated image.” ↑
  6. 6Midjourney, Upscalers, read at source 13 Oct 2026“If you choose the Creative upscaler, it will change your image a bit as it adds new details. On the other hand, the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible.” ↑a“the Subtle upscaler keeps changes to a minimum, preserving the look of your initial image as closely as possible.” ↑b
  7. 7CodeFormer, CodeFormer README (GitHub), read at source 13 Oct 2026“Fidelity weight w lays in [0, 1]. Generally, smaller w tends to produce a higher-quality result, while larger w yields a higher-fidelity result.” ↑
  8. 8OpenAI, Image prompting, read at source 13 Oct 2026“Repeated edits can still change details you intended to preserve. Restate those constraints and inspect each result.” ↑a“He has weathered skin with visible wrinkles, pores, and sun texture” and “The image should feel honest and unposed, with real skin texture, worn materials, and everyday detail. No glamorization, no heavy retouching.” ↑b“Request ‘photorealistic’ or ‘real photograph’ explicitly when that is the goal, and describe framing and texture. Treat camera specifications as cues for appearance, not a guarantee of exact physical simulation.” ↑c“Request ‘photorealistic’ or ‘real photograph’ explicitly when that is the goal, and describe framing and texture.” ↑d“State exclusions such as unwanted text, logos, or watermarks.” ↑e“No glamorization, no heavy retouching.” ↑f“Pass the previous output as the next edit input, request one change, and repeat the details to preserve. References such as ‘same style as before’ can carry context, but restate critical constraints if the result drifts.” ↑g“Request ‘photorealistic’ or ‘real photograph’ explicitly when that is the goal, and describe framing and texture.” ↑h
  9. 9Black Forest Labs, FLUX.2 prompting guide, read at source 13 Oct 2026“For photorealism, specify camera models, lenses, and film stocks.” ↑a“FLUX.2 does not support negative prompts. Focus on describing what you want, not what you don’t want.” and “Instead of ‘no blur,’ say ‘sharp focus throughout.’” ↑bAnalog Film “shot on Kodak Portra 400, natural grain, organic colors”; 80s Vintage “film grain, warm color cast, soft focus, 80s vintage photo”. ↑cquick reference table, “Guidance [flex] Prompt adherence guidance: 4.5 (1.5-10)”. ↑d“FLUX.2 does not support negative prompts.” ↑e
  10. 10OpenAI, Introducing ChatGPT Images 2.5, read at source 13 Oct 2026“GPT‑Image‑2.5 Flare brings the same improvements in quality, editing, and speed to the API”. ↑
  11. 11Google, Nano Banana image generation (Gemini API), last updated 9 Oct 2026, read at source 13 Oct 2026“A photorealistic [type of shot] of a [subject description] in a [setting description]. [Description of the light]. Shot from a [camera angle] with a [lens type].” ↑a“Use ‘semantic negative prompts’: Instead of saying ‘no cars,’ describe the intended scene positively: ‘an empty, deserted street with no signs of traffic.’” ↑b“Use the conversational nature of the model to make small changes. Follow up with prompts like, ‘That’s great, but can you make the lighting a bit warmer?’ or ‘Keep everything the same, but change the character’s expression to be more serious.’” ↑c
  12. 12Midjourney, Raw, read at source 13 Oct 2026“When you switch to Raw, you’re essentially turning off this ‘auto-pilot.’ With simple prompts, you’ll often get more realistic, photo-like images.” and “Raw is compatible with Midjourney versions 5.1 and later.” ↑a“Raw is compatible with Midjourney versions 5.1 and later.” and “Add --raw to the end of your prompt in the Imagine bar.” and “You can also turn on Raw for all your images in the settings panel.” ↑b
  13. 13Midjourney, No, read at source 13 Oct 2026“Midjourney’s moderation system reads every word you add to the --no parameter independently. This means if you prompt ‘--no modern clothing’ it will read that as ‘no modern’ and ‘no clothing’!” ↑
  14. 14Hugging Face, Diffusers: Text-to-image (StableDiffusionPipeline), read at source 13 Oct 2026negative_prompt “The prompt or prompts to guide what to not include in image generation.”; guidance_scale “defaults to 7.5”; “A higher guidance scale value encourages the model to generate images closely linked to the text prompt at the expense of lower image quality.” ↑anegative_prompt “Ignored when not using guidance (guidance_scale < 1).” ↑b
  15. 15Hugging Face, Diffusers: Stable Diffusion XL, read at source 13 Oct 2026guidance_scale “defaults to 5.0”; “Higher guidance scale encourages to generate images that are closely linked to the text prompt, usually at the expense of lower image quality.” ↑
  16. 16Hugging Face, Diffusers: Image-to-image, read at source 13 Oct 2026“a higher strength value gives the model more ‘creativity’ to generate an image that’s different from the initial image” and “a lower strength value means the generated image is more similar to the initial image”. ↑

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE