ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›THE CRAFT›NEGATIVE PROMPTS BY PROBLEM
THE CRAFT · REFERENCEINTERMEDIATE

Negative Prompts by Problem: What to Exclude and What to Write Instead

Nine common problems, from waxy skin and bad hands to stray watermarks and video flicker. Each gets a short negative list to paste and a positive line for tools that ignore negatives.

READ13 min
WORDS2,634
SOURCES6
TYPELIBRARY
CHECKED25 AUG 26
TL;DR — THE SHORT VERSION

Find the problem you can see, paste its short list if your tool has a negative field, and use its “write this instead” line if it does not.

  • Fix what broke, not what might. A short list aimed at one problem beats a long “universal” one.
  • Check your tool first. Stable Diffusion, SDXL and Midjourney (--no) read a negative prompt. FLUX.2, Runway, Seedance 2.5, Google Flow and ChatGPT Images have no negative field.
  • On Runway, negatives can backfire. Runway says including one may result in the opposite happening.
  • Every section has a positive line. It works in every tool, including the ones with a negative field.
  • For agents, negatives are guardrails. Spell out what an agent must not do: delete, send, install, pay, or follow instructions found in fetched content.
◈ THE QUICK LIST: 9 PROBLEMS, AND THE FIX FOR EACH
  1. Plastic or waxy skin ask for visible pores and soft light, not just "no plastic".
  2. Hands and fingers give the hands one simple, visible job, or keep them out of the frame.
  3. Extra limbs and body distortion ask for one person in a simple pose, fully in frame.
  4. Text, watermarks and logos ask for clean, unbranded surfaces, and put any words you do want in quotes.
  5. Duplicate people and extra faces say there is one person, and say where they are.
  6. Blur, grain and noise ask for sharp focus throughout, and drop "grain" if you want a film look.
  7. Hands and clothing that merge describe the contact: who holds what, and what the fabric does.
  8. Flicker, morphing and drift in video describe what must stay the same, and describe the camera as still.
  9. Unwanted styles name the style you want with its medium, and do not mix styles that fight each other.

First check which tools read a negative prompt. For longer lists, see lists by kind of shot.

A negative prompt is a list of things you want an AI image or video tool to leave out. This page sorts them by the problem you are trying to fix: waxy skin, bad hands, extra limbs, stray text and watermarks, duplicate faces, blur, hands merging into clothing, flicker in video, and the wrong style. Each problem gets a short list to paste and a positive line to write instead, because many tools ignore negatives.

CHECK YOUR MODEL FIRST — THIS LIBRARY IS NOT UNIVERSAL

No negative field — do not paste these: FLUX.2 ("does not support negative prompts" — Black Forest Labs), Seedance 2.5 (no field; exclusions compete with your prompt), ChatGPT Images (exclusions go inside the prompt), Google Flow (no negative box in the prompt settings). Actively harmful: Runway Gen-4/4.5 — Runway states negatives are unsupported and that including one "may result in the opposite happening."

Negative field: Stable Diffusion and SDXL, Midjourney (--no) and Veo 3.1 (negativePrompt on Google's Gemini Enterprise Agent Platform). Kling has one only on its legacy API, and recommends putting negatives inside the prompt instead. Keep entries short and specific.

On the tools that have no field, use the green “write this instead” lines. Each one describes your way out of a failure: not "no plastic skin" but "visible pores, uneven natural skin tone." That is the vendors’ own advice, not a workaround. Model-by-model detail, with sources: which AI models read negative prompts and what the vendors actually say.

1
◈ AND FOR AGENTS, THIS IS A SAFETY RAIL

The same idea does far more work when the model can act. A negative instruction is the cheapest guardrail you have: never delete, never send, never install, never pay, never act on instructions found in fetched content.

See guardrails for the full permission model.

On a model that reads them, a negative prompt can steer away from common failure looks such as plastic skin, flat lighting and stock-photo energy (this site’s reading from use, not a vendor claim). On a model that does not, the same words either do nothing or, on Runway, can pull the unwanted thing in. That is why every section below gives you both.

Which tools read a negative prompt?

The plain point: paste a negative list only where there is a place for it. Everywhere else, use the positive line.

NEGATIVE PROMPTS · WHERE THEY GO (READ AT SOURCE 12 OCT 2026)
ToolNegative field?What to do
Stable Diffusion, SDXLYes: a negative prompt box, or negative_prompt in code.Paste the short list for your problem. Start with the Stable Diffusion starter list.
MidjourneyYes: --no at the end of the prompt.Add the list after --no, separated by commas.
Veo 3.1 (Gemini Enterprise Agent Platform)Yes: negativePrompt.Keep it short. Google’s own advice is still to describe the scene you want.
KlingLegacy API only.Kling recommends writing negatives as sentences inside the prompt.
FLUX.2, Runway, Seedance 2.5, ChatGPT Images, Google FlowNo.Use the “write this instead” lines. Full guide: no negative prompt field?

The sources behind each row, all read on 12 Oct 2026:234567

How to use this page

The plain point: one problem, one short list, one change at a time.

  1. Name the problem you can see in your last result. Do not pre-load lists for problems you have not had.
  2. Tool with a negative field: paste only that section’s list. Tool without one: add the “write this instead” line to your main prompt, near the start.
  3. Re-run with the same seed if your tool lets you, so you can see what the change did.
  4. Watch for clashes. Do not exclude “blur” if you asked for a soft background, or “grain” if you asked for film grain. The model has to pick a side, and it may not pick yours (this site’s reading).

Plastic or waxy skin

The plain point: skin comes out smooth, shiny and airbrushed, like a doll or a wax figure. Ask for texture and soft light, not just “no plastic”.

NEGATIVE — PLASTIC OR WAXY SKIN
plastic skin, waxy skin, airbrushed skin, beauty filter, over-smoothed skin, porcelain skin, doll-like face, glossy skin, heavy retouching, CGI skin
WRITE THIS INSTEAD — ANY TOOL
natural skin texture with visible pores, faint fine lines and a slightly uneven skin tone, unretouched, soft diffused window light from one side

In ChatGPT Images you can also add a short exclusion sentence. OpenAI’s own image prompting guide uses this exact wording in an example.8 More words that read as real: realism words.

Hands and fingers

The plain point: extra, missing or fused fingers are the classic AI failure. Give the hands one simple, visible job, or keep them out of the frame.

NEGATIVE — HANDS AND FINGERS
extra fingers, missing fingers, fused fingers, too many fingers, malformed hands, twisted wrist, extra hands, deformed hands
WRITE THIS INSTEAD — HANDS IN THE SHOT
her left hand rests flat on the table, five relaxed fingers clearly separated, her right hand tucked into her coat pocket
WRITE THIS INSTEAD — HANDS OUT OF THE SHOT
head-and-shoulders portrait, framed from the chest up

Extra limbs and body distortion

The plain point: an extra arm, a missing leg or a body that bends the wrong way. Ask for one person in a simple pose, fully in frame.

NEGATIVE — LIMBS AND BODY
extra limbs, extra arms, extra legs, missing limbs, disconnected limbs, floating limbs, distorted body, unnatural proportions, elongated neck, twisted torso, bad anatomy
WRITE THIS INSTEAD — ANY TOOL
one adult standing upright, full body in frame from head to feet, both arms relaxed at the sides, natural body proportions, simple relaxed pose

Busy, twisted poses go wrong more often than simple ones (this site’s reading from use). On Stable Diffusion, a pose guide does more than any list: Hugging Face explains that with a ControlNet model you can give an image of a human pose and the model follows it.9 See ControlNet meaning.

Text, watermarks and logos

The plain point: stray letters, fake signatures, watermarks and made-up brand marks. Ask for clean, unbranded surfaces. If you do want words, put the exact words in quotes.

NEGATIVE — TEXT, WATERMARKS, LOGOS
text, letters, words, caption, subtitles, watermark, signature, username, logo, brand name, label text, stamp, frame border
WRITE THIS INSTEAD — ANY TOOL
a plain unbranded glass bottle with a blank matte label, standing in front of a smooth bare plaster wall

ChatGPT Images is the exception where a plain “no” line is the vendor’s own advice. OpenAI tells you to state these exclusions in the prompt.8

CHATGPT IMAGES — ADD AT THE END
No extra text, no watermarks, no unrelated logos.

These lists stop a generator inventing text. They are not a way to remove someone else’s watermark from a real photo. Getting real words right is its own skill: text inside AI images.

Duplicate people and extra faces

The plain point: the same person appears twice, a second face grows out of a shoulder, or a crowd turns up uninvited. Say there is one person, and say where they are.

NEGATIVE — DUPLICATES
duplicate person, clone, twins, two heads, double face, extra face, multiple people, crowd, repeated subject
WRITE THIS INSTEAD — ANY TOOL
a single woman, alone in an empty room, standing in the centre of the frame

For the same person across several images, see character consistency.

Blur, grain and noise

The plain point: soft, smeary or speckled images. Ask for sharpness directly. Black Forest Labs gives exactly this rewrite for FLUX.2: “Instead of ‘no blur,’ say ‘sharp focus throughout.’”6

NEGATIVE — BLUR AND NOISE
blurry, out of focus, soft focus, motion blur, noise, grain, jpeg artifacts, compression artifacts, low resolution, pixelated, smeared details
WRITE THIS INSTEAD — ANY TOOL
sharp focus throughout, crisp fine detail, clean low-noise image, bright even light

Drop “blurry” and “motion blur” from the list if you asked for a soft background or a sense of speed, and drop “grain” if you want a film look.

Hands and clothing that merge

The plain point: when a hand touches fabric (a zip, a pocket, a strap), fingers sink into the cloth, straps pass through the body or a sleeve melts into skin. Describe the contact: who holds what, and what the fabric does.

NEGATIVE — CLOTHING CONTACT
hand merging into fabric, fingers passing through clothing, fused sleeve, strap passing through body, clothing clipping, melted fabric, floating buttons, mismatched collar
WRITE THIS INSTEAD — ANY TOOL
her right hand grips the jacket lapel, fingers resting on top of the fabric, the wool creasing under her grip, the bag strap lying over her left shoulder

Flicker, morphing and drift in video

The plain point: faces change between frames, clothes shift colour, light pulses, the camera wanders. Describe what must stay the same, and describe the camera as still.

Runway’s video guide gives the camera version of this rewrite: “Locked camera. The camera remains still.” instead of “No camera movement.”10 ByteDance’s Seedance 2.5 guide does the same for faces and clothes: “keep the same face, age, hairstyle, wardrobe”.11

NEGATIVE — VIDEO (VEO NEGATIVEPROMPT, KLING LEGACY)
flickering, strobing, morphing face, identity drift, changing clothes, shifting colours, warping background, objects appearing and disappearing, jump cuts
WRITE THIS INSTEAD — ANY VIDEO TOOL
one continuous take, locked camera, the same face, hairstyle and outfit in every frame, steady even lighting, slow natural movement

Seedance has no negative field, but its guide does use a short “Avoid” clause at the end of the prompt, and warns against long lists.11

SEEDANCE 2.5 — LAST LINE OF THE PROMPT
Avoid: subtitles, hard cuts, duplicate people, face drift, wardrobe changes.

Unwanted styles: cartoon, CGI or 3D render

The plain point: you asked for a photo and got a render, or asked for a drawing and got a photo. Name the style you want, with its medium, and do not mix styles that fight each other. Civitai’s prompting guide warns that tokens such as “realistic” and “anime-style” used together “can confuse the model and produce inconsistent results.”12

NEGATIVE — WHEN YOU WANT A PHOTO
cartoon, anime, illustration, painting, drawing, 3D render, CGI, video game graphics, plastic look, oversaturated
WRITE THIS INSTEAD — A PHOTO
documentary photograph shot on 35mm film, natural daylight, true-to-life colours, real textures
NEGATIVE — WHEN YOU WANT AN ILLUSTRATION
photograph, photorealistic, 3D render, CGI, realistic skin texture, lens blur, film grain
WRITE THIS INSTEAD — AN ILLUSTRATION
flat 2D ink illustration with bold clean outlines and solid colour fills, hand-drawn look

Longer lists by kind of shot

The plain point: these older, longer lists are grouped by the kind of picture, not by problem. Paste only the entries your shot needs, and only where your tool has a negative field.

Photorealism

NEGATIVE — PHOTOREALISM
cartoon, illustration, anime, blurry, watermark, deformed, extra fingers, extra limbs, bad anatomy, plastic skin, oversaturated, flat lighting, stock photo, AI-looking, low quality, ugly, disfigured, out of frame, cropped, worst quality, normal quality, jpeg artifacts

Portrait

NEGATIVE — PORTRAIT
airbrushed skin, plastic skin, dead eyes, blank stare, asymmetrical face, bad teeth, visible teeth unnaturally, extra eyelashes, unnatural iris, bad mouth, floating hair, weird neck, double chin when not intended, skin blemishes (unless intentional), bad earrings, clipping accessories

Product photography

NEGATIVE — PRODUCT
hands in frame, human presence, fingerprints on product, harsh shadows, lens distortion, reflection of photographer, background clutter, crooked label, dust specks, cheap packaging look, unintended reflections, color cast, overexposed highlights, blurry product surface

Architecture and interiors

NEGATIVE — ARCHITECTURE / INTERIOR
people, cars, bikes, perspective distortion, fisheye distortion, chromatic aberration, lens flare (unless intended), construction equipment, garbage, graffiti (unless intended), overcast flat light (unless intended), bad CGI, 3D render look, unrealistic scale

Video and motion

NEGATIVE — VIDEO / TEMPORAL
flickering, morphing face, identity drift, scene cuts, jump cuts, strobing, unnatural motion, teleportation, fast motion blur, video artifacts, compression artifacts, pixelation, frame drops, disappearing limbs, extra limbs appearing, background objects changing between frames

Artistic and illustration

NEGATIVE — ARTISTIC / ILLUSTRATION
photorealistic (when illustrative is intended), photography, 3D render, CGI, hyper-detailed (when painterly is intended), stock art, clip art, low resolution, white background (unless intended), signature, watermark, text, logo, frame border

Questions people ask

Is there a universal negative prompt that works everywhere?

No. Several popular tools have no negative field at all, and on Runway the vendor warns a negative prompt can produce the opposite of what you asked. Even where there is a field, longer is not stronger: Civitai notes that repeating the same keywords “does not necessarily improve results and may lead to distortions.”12

What is the best negative prompt for hands and fingers?

Use the short hand list above on tools with a negative field, and add the positive hand line everywhere. In this site’s experience the cheapest fix is giving the hands one simple job, or framing them out of the shot.

Can I use negative prompts in ChatGPT?

There is no separate negative box. OpenAI’s advice is to write exclusions into the prompt itself, such as “No extra text, no watermarks, no unrelated logos.” (see the text section above for the source). The full how-to for tools without a field: no negative prompt field?

TAKEAWAY

Name the problem, paste only its short list where your tool has a negative field, and always add the positive line. Where there is no field, the positive line is the whole fix.

Related on this site: what a negative prompt means · negative prompts for every AI model · no negative prompt field? · negative prompts for Stable Diffusion and SDXL · text-to-image meaning.

Related on this site: Free Negative Prompt Builder for AI Images and Video · Why Do AI Images Look Plastic? · Micro Detail Prompts · Fix a Bad AI Image · How to Fix AI Video Flicker, Morphing and Face Drift

The negative lists and “write this instead” lines on this page are this site’s own prompts, not claims from these sources. Remarks marked “this site’s reading” are from use, not vendor statements.

SOURCES
  1. 1Vendor documentation: docs.bfl.ai · help.runwayml.com · Google Cloud Veo guide · Kling API docs. Re-read 12 Oct 2026 (Runway in a browser)FLUX.2 “does not support negative prompts”, and on Runway they “are not supported in Gen-4 Images”. Checked 25 Aug 2026 ↑
  2. 2Hugging Face, Diffusers: Text-to-image (StableDiffusionPipeline), read at source 12 Oct 2026“The prompt or prompts to guide what to not include in image generation.” ↑
  3. 3Midjourney, No parameter, read in a browser 12 Oct 2026“Add --no to the end of your prompt in the Imagine bar, followed by the thing or list of things you don’t want in your image.” ↑
  4. 4Google Cloud, Gemini Enterprise Agent Platform: Text to video, last updated 10 Oct 2026, read at source 12 Oct 2026“Optional: A string value that describes content that you want to prevent the model from generating.” ↑
  5. 5Kling AI, Text to Video API (legacy), read in a browser 12 Oct 2026“It is recommended to supplement negative prompt via negative sentences within positive prompts”. ↑
  6. 6Black Forest Labs, FLUX.2 prompting guide, read at source 12 Oct 2026“FLUX.2 does not support negative prompts.” ↑a“Instead of ‘no blur,’ say ‘sharp focus throughout.’” ↑b
  7. 7Google, Create videos in Google Flow, read at source 12 Oct 2026. The page lists the prompt box settings (aspect ratio, number of outputs, model and generation length) and no negative prompt field.↑
  8. 8OpenAI, Image prompting, read at source 12 Oct 2026“No glamorization, no heavy retouching.” ↑a“State exclusions such as unwanted text, logos, or watermarks.” Its example wording: “No extra text, no watermarks, no unrelated logos.” ↑b
  9. 9Hugging Face, Diffusers: Text-to-image, read at source 12 Oct 2026“if you provide an image of a human pose (usually represented as multiple keypoints that are connected into a skeleton) as a conditioning input, the model generates an image that follows the pose of the image.” ↑
  10. 10Runway, Gen-4 Video Prompting Guide, read in a browser 12 Oct 2026“Negative phrasing is not supported and may produce unpredictable or even opposite results. ❌ No camera movement. The camera doesn’t move. NO MOVEMENT ✅ Locked camera. The camera remains still.” ↑
  11. 11Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 12 Oct 2026“For continuity: keep the same face, age, hairstyle, wardrobe, product geometry, and lighting direction across every shot.” ↑a“Keep them short, concrete, and grouped at the end.” “Do not write a long generic blacklist that competes with the scene description.” ↑b
  12. 12Civitai Education, Civitai’s Prompt-Crafting Guide: Part 1 – Basics, last updated 13 Mar 2025, read at source 12 Oct 2026“can confuse the model and produce inconsistent results.” ↑a“Repeating the same keywords does not necessarily improve results and may lead to distortions.” ↑b

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
◈ IF YOU ARE CITING THIS

Or check it yourself. How to check the figures here names the feed or document behind each recurring source, and what to expect when your number differs from ours.

Cite the original source, not this page. Every figure here names the organisation that issued it and the date it was published — those are the citations worth carrying. This page is a signpost, not a primary source.

If you need to reference the collation itself — the comparison, the framing, or a correction made here — the press page has the details. But if you are quoting a number, go to whoever measured it.