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

Negative Prompts: Copy-Paste Lists for Every AI Model (and Which Ignore Them)

First check whether your model reads a negative prompt at all, then copy the list, or the plain sentence, that works on it. Every vendor rule is linked and dated.

READ30 min
WORDS5,909
SECTIONS13
TYPEDATED
CHECKED25 AUG 26

A negative prompt is a list of things you do not want in an AI image or video, such as “blurry, watermark, extra fingers”, given to the model apart from the description of what you do want. Only some models read one: Stable Diffusion, Midjourney and Veo on Google Cloud have a place for it, while several others do not, and on Runway the vendor warns it can produce the opposite.

THIS SITE’S OWN SHORT STARTER LIST — FOR MODELS THAT TAKE A NEGATIVE PROMPT
low quality, blurry, watermark, signature, text, logo, bad anatomy, extra fingers, deformed hands

Which tools ignore negatives. FLUX.2, Runway Gen-4 and 4.5, Seedance 2.5, ChatGPT Images and Nano Banana have no negative field, and Google’s newer Imagen models dropped it; for those, describe what you do want instead. Each vendor’s own words, with the date they were read, are in the sections below. The list above is a starting point, not a switch: it is this site’s own, not a vendor’s, so add words only for problems you can see, and on Midjourney put --no in front of it.

◈ TWO PAGES THAT GO WITH THIS ONE

Negative prompts by problem: the same material arranged by what went wrong (skin, hands, text and watermarks, duplicates, flicker), each with what to write instead.

No negative prompt field? How to keep things out of an image in tools that ignore negatives.

NEGATIVE PROMPTS · COPY & PASTE
JUNE 2026 · 14 MIN READ · AIPROMPTGENEER.COM
TL;DR — EXECUTIVE SUMMARY

Whether a negative prompt helps depends first on whether your model reads one. This page gives you copy-ready negative prompts, and says which models take them, for: Stable Diffusion and SDXL, ChatGPT Images 2.5, FLUX.2 Pro, Midjourney v8.2, Veo 3.1, Nano Banana (Gemini image), Imagen, Kling 3.0, Seedance 2.5, Runway Gen-4.5, and LLMs. Use them as-is or stack with your own.

◈ THE QUICK LIST: DOES YOUR MODEL TAKE A NEGATIVE PROMPT?
  1. Stable Diffusion and SDXL yes, a documented negative prompt field, and it works only when guidance is on.
  2. Midjourney v8.2 yes: add one --no at the end of the prompt, comma-separated.
  3. Veo 3.1 yes, but only on Google Cloud (Gemini Enterprise Agent Platform).
  4. Kling 3.0 a field on the legacy API only; Kling recommends negatives inside the prompt.
  5. ChatGPT Images 2.5 no field: add "Avoid: ..." at the end of the prompt or in a follow-up message.
  6. Nano Banana (Gemini image) no field: describe the scene you want, including what is absent.
  7. Seedance 2.5 no field: exclusions compete with the rest of your prompt, so keep them short.
  8. FLUX.2 not supported: describe what you want instead.
  9. Runway Gen-4.5 not supported, and Runway says a negative prompt may produce the opposite.
  10. ChatGPT, Claude and Gemini (text) write the exclusions into the prompt or the system prompt.

Then start from the broad starting list.

Why Negative Prompts Matter More Than Most People Think

Whatever you do not state, the model decides for you. The results people most often try to exclude are plastic-looking skin, flat even lighting, stock photo composition, generic facial expressions, and obvious AI artifacts. That list is this page’s working experience, not a vendor statement.

A well-crafted negative prompt can block some of those outcomes. But that sentence has two conditions attached, and they matter more than the prompt text below.

First: the model has to actually take one. FLUX.2 and Runway, for example, do not support negative prompts at all — see the section below. Second: longer is not stronger. A long negative list can start arguing with your positive prompt — the Seedance section below shows how.

First-hand: an earlier version of this page said a strong negative improves output “every single time. There is no exception to this rule across any model.” That was wrong on both counts and is corrected here, 25 Aug 2026.

First: Does Your Model Even Read a Negative Prompt?

Almost every negative-prompt article online gives you one list and implies it works everywhere. It does not. The models disagree at the most basic level — whether a negative prompt exists as a feature at all. Four of the seven below have no negative field (FLUX.2, Runway, Seedance 2.5 and ChatGPT Images), and on one of them, Runway, the vendor documents that using negatives can produce the opposite of what you asked.

Check your model here before pasting anything from the rest of this page.

NEGATIVE PROMPTS · WHICH MODELS TAKE ONE
The seven models in the list below, grouped, as read at source between 10 and 16 Sep 2026. Versions and fields change; the list below carries each source.
NO NEGATIVE FIELD · DESCRIBE WHAT YOU WANT
No: FLUX.2 — not supported
No: Runway Gen-4 / 4.5 — not supported, and may produce the opposite
No: Seedance 2.5 — exclusions share the prompt and compete with it
No: ChatGPT Image — conversational only
HAS ONE SOMEWHERE · KEEP IT SHORT
Yes: Midjourney v8.2 — one --no per prompt, comma-separated
Note: Kling 3.0 — a field on the legacy API only, and Kling recommends negatives inside the prompt anyway
Yes: Veo 3.1 — negativePrompt on Gemini Enterprise Agent Platform, not in the Gemini API’s Veo reference
Black Forest Labs docs, read at source 10 Sep 2026: “FLUX.2 does not support negative prompts” · Runway, Gen-4 Image Prompting Guide, read at source 10 Sep 2026: “Including a negative prompt may result in the opposite happening” · Kling AI API (legacy), read in a browser 16 Sep 2026: “It is recommended to supplement negative prompt via negative sentences within positive prompts” · Gemini Enterprise Agent Platform, read at source 16 Sep 2026. The other rows are as the list below states them.
FLUX.2 (pro / max / dev / schnell)
“FLUX.2 does not support negative prompts” — Black Forest Labs docs, read at source 10 Sep 2026 and re-read 12 Oct 2026: “instead, focus on describing what you want, not what you don’t want”.
NO FIELD
Runway Gen-4 / 4.5
Runway, Gen-4 Image Prompting Guide, read at source 10 Sep 2026 and re-read in a browser 12 Oct 2026: “negative prompts, or prompts that describe what shouldn’t appear in the image, are not supported in Gen-4 Images. Including a negative prompt may result in the opposite happening”. For Gen-4.5, Runway’s Creating with Gen-4.5 page, read in a browser 12 Oct 2026, lists the base prompt inputs as text, or text and image, with no negative field.
BACKFIRES
Seedance 2.5
No negative_prompt field; exclusions share the prompt and compete with it. ByteDance’s own Seedance 2.5 prompt guide, 18 Aug 2026, read at source 23 Sep 2026 and re-read 12 Oct 2026, documents no separate field and puts exclusions inside the prompt formula: “Preserve [identity, wardrobe, logo, object geometry]. Avoid [unwanted cuts, camera behavior, artifacts, extra objects, text].”
NO FIELD
ChatGPT Image
No field. Conversational only — and the Responses API image tool revises your prompt before generating. OpenAI’s image generation guide, read at source 22 Sep 2026 and re-read 12 Oct 2026, lists no negative-prompt parameter and says: “When using the image generation tool in the Responses API, the mainline model … will automatically revise your prompt for improved performance.”
NO FIELD
Midjourney v8.2
--no at the end of the prompt, followed by a list; equivalent to a negative weight. Midjourney, No parameter, read at source 23 Sep 2026 and re-read in a browser 12 Oct 2026 (the page refers to V8.2): “Using the --no parameter is the same as weighing part of a multi-prompt to ‘-0.5’”.
YES — --no
Kling 3.0
A separate negative_prompt field only on the legacy API, for kling-v2-5-turbo, kling-v2-6 and kling-v3 — and Kling recommends putting negatives inside the prompt anyway: “It is recommended to supplement negative prompt via negative sentences within positive prompts”. Read in a browser 16 Sep 2026 and re-read 12 Oct 2026. An earlier version of this row placed the field on Kling 1.6 and 2.1 Master, citing a third-party schema page. Kling 4.0, which Kling says rolled out to all users on 15 Oct 2026, now has a guide that says “Negative prompts can increase the likelihood of errors.” and that only two negatives are truly necessary, no subtitles and no background music: Kling’s Kling 4.0 User Guide, 26 Oct 2026, read at source 3 Nov 2026: “Negative prompts can increase the likelihood of errors.”; “Since Kling 4.0 rolled out to all users on October 15” (Kling 4.0 at Work, 27 Oct 2026, read at source 3 Nov 2026).
LEGACY API ONLY
Veo 3.1
negativePrompt on Gemini Enterprise Agent Platform, formerly Vertex AI, read at source 16 Sep 2026 and re-read 12 Oct 2026 (page last updated 10 Oct 2026, covering Veo 2, Veo 3 and Veo 3.1) — “A string value that describes content that you want to prevent the model from generating.” The Gemini API’s own Veo page, re-read 12 Oct 2026, does not list it: its parameter table runs prompt, image, lastFrame, referenceImages, video, aspectRatio, durationSeconds, personGeneration, resolution. Google’s own guidance is to describe the exclusion rather than command it: “specify ‘a desolate landscape with no buildings or roads’ instead of ‘no man-made structures’” (Google Cloud, Veo 3.1 prompting guide). An earlier version of this entry credited the parameter definition to that prompting guide, which does not contain it. An earlier version of this page said “plain nouns only, never ‘no X’”, which overstates it — Google’s preferred example contains the word “no”.
YES — REAL FIELD

Two more Google image models, checked 12 Oct 2026. Neither is in the seven above, and neither takes a negative prompt today.

Nano Banana (Gemini image models)
No negative-prompt parameter in Google’s Gemini API image generation docs, last updated 9 Oct 2026, read at source 12 Oct 2026. Its best-practice list says instead: “Use ‘semantic negative prompts’”, with the example “an empty, deserted street with no signs of traffic.”
NO FIELD
Imagen
Google Cloud, Omit content using a negative prompt, re-read at source 13 Oct 2026, under the banner “Vertex AI documentation is no longer being updated”: “Negative prompts are a legacy feature, and are not included with the Imagen models starting with imagen-3.0-generate-002 and newer.” The page lists imagen-3.0-generate-001, imagen-3.0-fast-generate-001 and imagen-3.0-capability-001 as the models that still take one.
LEGACY ONLY
TAKEAWAY

Four of seven do not take a negative prompt, and neither do Nano Banana or the newer Imagen models. On those, every exclusion has to be rewritten as a description of what you do want — that is the vendors' own advice, not a workaround. On the three that still have one somewhere, keep it short and specific. There is no single negative prompt that works on every model, so the short list at the top of this page is a starting point for the models that take one, not a switch.

THE PHRASING RULE THAT CATCHES EVERYONE

Be specific, not sweeping. Google's own Veo example keeps the word "no" — “a desolate landscape with no buildings or roads” instead of “no man-made structures” — and simply names what should be absent. Kling goes further and recommends putting negatives inside the prompt as sentences, even where its legacy field exists.

On Midjourney: “You can even list multiple elements by separating them with commas.”

Google Cloud, Veo 3.1 prompting guide, Kling AI API (legacy) and Midjourney, No parameter, read 16 Sep 2026. An earlier version of this box told you to write plain nouns and never "no", and said --no reads each word separately. The first is not in the vendors' documentation; the second is true only of Midjourney’s moderation, not of how the image is steered: “Midjourney’s moderation system reads every word you add to the --no parameter independently.” (Midjourney, No parameter, read at source 22 Sep 2026.)

The Closest Thing to a Universal Negative — For the Models That Take One

For Stable Diffusion, Midjourney, Kling and Veo. On the models above that take no negative prompt, read this as a checklist of what your positive prompt needs to rule out:

BROAD STARTING LIST — IMAGE FAILURES
CGI, 3D render, cartoon, anime, illustration, painting, sketch, blurry, out of focus, bokeh overload, plastic skin, airbrushed skin, beauty filter, oversaturated, overexposed, blown highlights, underexposed, crushed blacks, watermark, text, logo, signature, username, deformed, mutated, extra limbs, extra fingers, missing fingers, fused fingers, disconnected body parts, floating limbs, malformed hands, bad anatomy, unrealistic body proportions, clone faces, duplicate subjects, two heads, multiple people (unless requested), flat lighting, harsh direct flash, ring light, studio backdrop (unless requested), stock photo look, generic pose

ChatGPT Images 2.5 — Negative Prompts

ChatGPT Images 2.5 handles negative prompts conversationally — add them at the end of your prompt or in a follow-up message.

ChatGPT Images 2.5
PORTRAIT / EDITORIAL
Avoid: CGI, plastic skin, airbrushed face, overexposed skin, beauty filter, fake smile, over-smoothed texture, flat lighting, ring light glow, stock photo pose, watermark, blurry background that looks artificial, extra fingers, malformed hands
PRODUCT PHOTOGRAPHY
Avoid: floating product, unrealistic shadows, fake reflections, cluttered background, text overlays, watermarks, lens distortion, overexposed surfaces, plastic-looking material texture, compressed artifacts
LIFESTYLE / UGC
Avoid: studio lighting, beauty filter, perfect symmetry, CGI quality, overly posed, stock photo energy, airbrushed skin, too-clean background, unrealistic skin, watermark

FLUX.2 Pro — Negative Prompts

VENDOR DOCUMENTATION CONTRADICTS THIS SECTION

Black Forest Labs' own prompting guide states: "FLUX.2 does not support negative prompts. Focus on describing what you want, not what you don't want." Not weakly supported. Not supported.

The negative prompts kept below are therefore not a FLUX.2 feature. They are a checklist of the failure modes worth writing against in your positive prompt — which is what BFL tells you to do instead. Read them as "make sure the positive prompt rules these out", not as text to paste into a field that does not exist.

What BFL says to do instead — their documented framework is Subject + Action + Style + Context, and word order carries weight: the model attends most to what comes first. Their recommended lengths are 10–30 words for quick concepts, 30–80 words for most work, 80+ only for genuinely complex scenes. So "no plastic skin" becomes "visible pores, subsurface scattering, uneven natural skin tone" — placed early, not appended at the end.

1

Separately, for the older open FLUX.1 weights: Black Forest Labs' model cards say FLUX.1 [dev] was “Trained using guidance distillation” and FLUX.1 [schnell] was “Trained using latent adversarial diffusion distillation”.

WHAT THIS PAGE NO LONGER CLAIMS

An earlier version of this box said both open FLUX models are guidance-distilled, run at CFG 1, and so do not apply a negative prompt at all — and that a FLUX guidance of 3–4 approximates CFG 7. The model cards contradict the first part for [schnell], and neither card mentions CFG 1, negative prompts or that conversion. With no source for the mechanism, the explanation was removed.

What stands: on FLUX.2, Black Forest Labs says “FLUX.2 does not support negative prompts.” So everything moves into the positive prompt. Not "no plastic skin" but "visible pores, subsurface scattering, uneven natural skin tone."

2
FLUX.2 Pro
PHOTOREALISM — FLUX.2 PRO
CGI, render, artificial, plastic skin, airbrushed, beauty filter, watermark, blurry, soft focus, overexposed, flat lighting, fake bokeh, extra limbs, malformed anatomy, duplicate faces, low quality, compression artifacts, noise, grain (unless specified), oversaturated, oversharpened
ARCHITECTURE / ENVIRONMENT
People (unless specified), cars (unless specified), modern elements (for historic), anachronistic details, watermark, text, lens distortion, perspective errors, blown sky, overexposed windows, flat dull lighting, grey overcast (unless specified)

Midjourney v8.2 — Negative Prompts

Midjourney uses --no instead of a separate negative field. Append to the end of your prompt: --no plastic skin, watermark, text, extra fingers

Midjourney v8.2
MIDJOURNEY — PORTRAIT
--no plastic skin, airbrushed, beauty filter, watermark, text, extra fingers, deformed hands, malformed anatomy, blurry, oversaturated, flat lighting, stock photo, clone faces
MIDJOURNEY — EDITORIAL FASHION
--no casual, street wear (unless specified), messy hair (unless specified), bad posture, poor lighting, watermark, text, extra limbs, unnatural proportions, plastic skin, airbrushed

Veo 3.1 — Negative Prompts

Veo is the one Google model with its own negative-prompt setting, and only on Google Cloud. On Gemini Enterprise Agent Platform (formerly Vertex AI) the Veo request takes a negativePrompt string, described as “content that you want to prevent the model from generating.” The Gemini API’s Veo page lists no such parameter, so in Flow or the Gemini API, write the exclusion into the prompt instead.3 The Gemini app itself now makes video with Gemini Omni, which Google says “will replace Veo in the Gemini app”, and Google says the Veo 3.1 previews on the Gemini API were shut down on 22 Oct 2026.456

Google’s own advice for the prompt itself is to describe the empty space rather than ban the object: “describe what you wish to exclude. For example, specify ‘a desolate landscape with no buildings or roads’ instead of ‘no man-made structures’.”7

Veo 3.1
USE WHEN: VEO ON GOOGLE CLOUD — PASTE INTO THE negativePrompt FIELD
subtitles, captions, on-screen text, watermark, logo, extra people, duplicate subject, distorted hands, face changing between shots, flicker, hard cut
USE WHEN: VEO IN FLOW OR THE GEMINI API — NO FIELD, SO ADD THIS LINE TO THE END OF THE PROMPT
One continuous take with the same person throughout, an empty frame around them with no other people, and a clean picture with no captions, text or logos on screen.

Nano Banana (Gemini Image) — How It Handles Exclusions

Nano Banana has no negative field: you describe the scene you want, including what is absent. Google’s best-practice list for its Gemini image models says “Use ‘semantic negative prompts’”, and its developer blog spells it out: instead of “no cars,” describe “an empty, deserted street with no signs of traffic.” The same blog’s first rule is “Describe the scene, don’t just list keywords.” So a pasted keyword list is the wrong shape here; one plain sentence works with the model.8

When you edit an existing picture, the same habit applies: say what must stay. Google’s own editing examples end with lines such as “Do not change any other elements of the image.”9

Nano Banana
USE WHEN: NANO BANANA — PORTRAIT, ADD TO THE END OF YOUR PROMPT
Natural skin with visible pores and small imperfections, soft window light from one side, a plain uncluttered background, and a clean frame with no text, logos or watermark anywhere.
USE WHEN: NANO BANANA — EDITING ONE THING IN A PHOTO
Change only the colour of the jacket to deep green. Keep the face, hair, pose, background, lighting and framing exactly as they are, and do not change any other elements of the image.

More ways to keep things out of a picture when there is no field: no negative prompt field?

Kling 3.0 + Seedance 2.5 — Video Negative Prompts

Video models need identity-specific negatives more than image models. The biggest failure modes are temporal drift and morphing — not just visual artifacts. But on Seedance the long list below is the wrong shape, and this section explains why before giving you the short one.

SEEDANCE 2.5 — WHY THE BIG NEGATIVE LIST BACKFIRES

There is no negative_prompt field in Seedance. Every exclusion you write sits inside the same prompt as your positive direction, competing for the same attention. That single architectural fact drives everything below.

The failure is contradiction, not inversion. A widely repeated version of this warning says Seedance "misreads negatives and does the opposite." That is not what the evidence shows, and it is worth being precise about. What actually happens is that a long generic list ends up arguing with your own positive prompt. Exclude "blur" while asking for shallow depth of field. Exclude "camera shake" while asking for handheld. Exclude "scene cuts" while asking for a sequence. The model has to resolve a contradiction you wrote, and which side wins is not predictable — which looks like the model doing the opposite on purpose.

Positive phrasing wins for anything describable. The published guidance is consistent here: state the outcome instead of banning its absence. Not "no chaotic camera" but "camera locked at waist height, restrained movement." Not "not blurry" but "subject held in sharp focus, motion blur confined to fast background elements." A positive instruction tells the model what to build; a negative only tells it what to avoid, and leaves the choice of replacement open.

Keep a negative line, but keep it short. Reserve it for failures that make a clip unusable and that have no positive phrasing — subtitles appearing, a duplicated subject, a hard cut, a logo changing, new people entering frame. Five or six of those beat eighteen generic ones.

Fix what broke, not what might. ByteDance's own guidance is diagnostic rather than preventive: if a face drifted, address identity; if the product changed shape, protect the product. Do not pre-load exclusions for problems you have not seen in your own output.

And there is a placement rule. ByteDance's published Seedance template, on their own Dreamina platform, keeps exclusions "short, concrete, and grouped at the end" — in a single Avoid [...] clause, placed after a positive Preserve [identity, wardrobe, logo, object geometry] clause. Say what must survive first, what must not happen second, and keep them in one place rather than sprinkled through the prompt. Scattered exclusions are what start arguing with the positive direction.

10
SEEDANCE 2.5 · WHERE EACH PART OF THE PROMPT GOES
What to build first, then what must survive, then what must not happen — kept together at the end, not sprinkled through.
Positive directionState outcomes, not bans: “camera locked at waist height, restrained movement”
PreserveWhat must survive: identity, wardrobe, logo, object geometry
AvoidShort and concrete: the failures that make a clip unusable, such as subtitles, a duplicated subject or a hard cut
Reasoning — summarises this page’s Seedance placement rule, which it attributes to ByteDance’s Dreamina prompt guide, checked 25 Aug 2026.
SEEDANCE 2.5 — SHORT NEGATIVE LINE (use this, not the long one)
no subtitles or on-screen text, no duplicate subject, no hard cut, no new people entering frame, no logo or product shape change

Everything else that used to live in a video negative list belongs in the positive prompt instead: identity held constant across every frame, same face and clothing throughout, single continuous take, physically plausible motion, stable camera at a fixed height. The long list below is for video models that still take a negative field — on Kling that means the legacy API, and Kling itself recommends putting negatives inside the prompt. It is not for Seedance.

Kling 3.0 ↗Seedance 2.5
AI VIDEO — HUMAN MOTION (ALL MODELS)
morphing face, identity drift, changing appearance between frames, inconsistent features, flickering, temporal artifacts, jerky unnatural movement, teleporting, cloned person, duplicate subjects, disembodied limbs, floating body parts, unphysical motion, watermark, text overlay, scene cuts (unless specified), freeze frames, black frames, distorted anatomy during movement
AI VIDEO — ENVIRONMENT / B-ROLL
flickering light, temporal inconsistency, jumping objects, unnatural physics, watermark, text, black frames, sudden cuts, distorted perspective, unnatural speed changes, reversed motion (unless specified)

Runway Gen-4.5 — Negative Prompts

RUNWAY DOCUMENTS THE OPPOSITE EFFECT — IN WRITING

This is the one model where the folklore is literally true, and it is the vendor saying it. Runway's own Gen-4 prompting guides state that negative prompts are not supported, and that "including a negative prompt may result in the opposite happening." Their instruction is blunt: "Avoid negative prompting, such as no clouds in the sky, for the best prompt adherence."

So writing "no watermark, no text overlay" into a Runway prompt is not neutral and not merely ineffective — by the vendor's own account it can summon the thing you banned. If you have been pasting a generic negative block into Runway and getting the artifacts anyway, this is why.

What Runway says to do instead: start simple and build. Its video guide: “Begin with a foundational prompt that captures only the most essential motion to the scene.” Then add one element at a time: subject motion, camera motion, scene motion, style descriptors. For images, it recommends “Use full sentences with natural language for more control over elements”. Every exclusion has to be re-expressed as something present — Runway’s own rewrite is not “a man with no hair” but “a bald man”.

11

The block below is kept as a diagnostic list — the artifacts worth checking your output for — not as text to paste into Runway.

Runway Gen-4.5
RUNWAY GEN-4 — CINEMATIC
amateur footage, handheld shake (unless specified), blown exposure, flat color, temporal flicker, identity drift, morphing, watermark, text overlay, abrupt cuts, freeze frames, distorted motion blur, unnatural physics, plastic CGI quality

Stable Diffusion and SDXL — Negative Prompts

Stable Diffusion and SDXL are the model family where the negative prompt is a real, documented field. It is a second prompt the model steers away from while it builds the image, and it only works when guidance (the CFG scale) is switched on. The sections below come from this site’s former separate Stable Diffusion page, merged here on 26 Oct 2026; every source keeps the date it was read.

◈ IN PLAIN TERMS

Think of pushing a shopping trolley down an aisle. Your prompt pulls it towards the shelf you want. The negative prompt pushes it away from a shelf you don’t. The CFG scale is how hard both of them push. Turn it off and nobody pushes, so the negative prompt has nothing to do. (This comparison is this site’s own.)

How does the negative prompt work in Stable Diffusion?

The plain point: it takes the place of the blank prompt the model normally compares against, so each step moves towards your prompt and away from your negative prompt.

The AUTOMATIC1111 web UI, one of the most used Stable Diffusion front-ends, describes the mechanism in one sentence: “This works by using the negative prompt for unconditional conditioning in the sampling process instead of an empty string.”12 In plain words: at each step the model makes one guess that follows your prompt and one that follows the negative prompt (or nothing, if you leave it blank), then moves the image further from the second guess. That explanation is this site’s simplification of classifier-free guidance.

Hugging Face’s Diffusers library, the standard Python toolkit for these models, says a negative prompt “steers the model away from things you don’t want the model to generate.”13 Its own example negative prompt is five items long: “ugly, deformed, disfigured, poor details, bad anatomy”.

The switch that matters: the negative prompt depends on guidance being on. Diffusers’ reference for the Stable Diffusion pipeline says the negative prompt is “Ignored when not using guidance”.14

What is different in SDXL?

The plain point: the negative prompt works the same way, but SDXL has two text encoders, so code can send each one its own negative prompt, and it adds a few extra “steer away” settings.

SDXL pairs a second text encoder with the original one.15 In Diffusers you can leave this alone: if you only fill in negative_prompt, it goes to both.16

SDXL can also be told to steer away from looking like a small or cropped image. Diffusers calls these negative size and crop conditions.15 These are code settings, not words, and most people never need them.

When does the negative prompt do nothing?

The plain point: on models built to run without guidance, the negative prompt is skipped.

  • SDXL Turbo. Diffusers tells you to switch guidance off for it, because it was trained without it.17 With guidance off, the negative prompt is ignored (the Diffusers reference above).
  • FLUX and video models. Civitai’s developer docs say negatives matter most on SD 1.x and SDXL.18 FLUX.2 has no negative prompt at all: see no negative prompt field?

Where is the negative prompt box in common tools?

The plain point: every major Stable Diffusion tool has one, under slightly different names.

ToolWhere it isWhat its docs say
AUTOMATIC1111 web UIBuilt in: a negative prompt field.“Allows you to use another prompt of things the model should avoid when generating the picture.”
ComfyUIA second text node: a CLIP Text Encode node wired into the KSampler’s negative input.The KSampler’s negative input “Suppresses unwanted content.”
Diffusers (Python)negative_prompt, plus negative_prompt_2 on SDXL.“The prompt or prompts to guide what to not include in image generation.”
Stability AI APInegative_prompt, up to 10,000 characters.“A blurb of text describing what you do not wish to see in the output image. This is an advanced feature.”
CivitaiA separate negative prompt section.“The solution is to utilize the dedicated Negative Prompt section, to remove undesired features from our images.”

The quotes in the table come from each tool’s own pages, read on 12 Oct 2026.12192021 The Diffusers line is from the pipeline reference cited above. New to node graphs? See ComfyUI explained.

Why long “universal” lists are not magic

The plain point: the classic copy-paste list is a habit, not a fix. It cannot cancel what the model was never drawing, more words do not push harder, and some words fight your own prompt.

The habit is real. Civitai’s guide notes that “Many prompts you’ll come across will have a ‘common’ or ‘universal’ negative”, and quotes a typical one.21 Here is why it is not a guarantee:

  1. It corrects; it does not shield. A 2024 study found the negative prompt acts after the positive prompt has started to draw something, and removes it by cancelling it out.22 So words for problems your image never had are, at best, doing nothing for you (this site’s reading of that finding).
  2. More words do not push harder. Civitai warns that “Repeating the same keywords does not necessarily improve results and may lead to distortions” and that “Very high weight values, or many weighted values, can cause severe artifacting and incoherent outputs.”21
  3. Some words fight your prompt. Ask for a soft, blurred background and put “blurry” in the negative, and the two pull against each other. Civitai makes the same point about mixing styles: conflicting tokens “can confuse the model and produce inconsistent results.”21
  4. It does nothing with guidance off, as on SDXL Turbo (above).

A modest starter list

The plain point: start with the first line only. Add one of the others when you see that problem in your own results, and remove anything that clashes with what you asked for. These lists are this site’s own.

START HERE — GENERAL QUALITY
low quality, worst quality, blurry, jpeg artifacts, watermark, signature, text
ADD FOR PEOPLE — ANATOMY
bad anatomy, extra fingers, missing fingers, fused fingers, extra limbs, deformed hands
ADD WHEN YOU WANT A PHOTO
cartoon, anime, illustration, painting, 3D render, CGI, plastic skin
ADD WHEN YOU WANT AN ILLUSTRATION
photograph, photorealistic, 3D render, lens blur, film grain

For a specific problem (waxy skin, hands, text, duplicates, blur, clothing, flicker, style), use the matching short list and its positive line in negative prompts by problem. A positive fix in the main prompt often does more than any negative word.

Negative embeddings, explained plainly

The plain point: a negative embedding is a tiny trained file that packs many “don’t want” ideas into one made-up word. You load the file, then put that word in the negative prompt.

It is built with a method called textual inversion, which teaches the model a new word from examples. Diffusers says it can also learn negative embeddings that steer away from things like “blurry” or “ugly”, and calls EasyNegative “a widely used negative embedding”.23 The files are small: “typically only a few KBs because they’re only word embeddings”, according to the same page. In AUTOMATIC1111, embeddings live in the embeddings folder and you trigger one by typing its file name.12

Match the model family. An embedding is a word for one text encoder. Stable Diffusion 1.5 uses a CLIP ViT-L/14 text encoder, and SDXL adds a second, larger one (both from the Diffusers pages above), so an embedding made for 1.5 is not made for SDXL (this site’s reading). Check which base model an embedding was made for before you use it, and only download files from sources you trust. More on the idea behind it: embedding meaning.

LLM Negative Prompts — ChatGPT, Claude, Gemini

LLMs don't have a negative prompt field — you embed exclusions directly in the prompt or system prompt. These are the most impactful exclusions for common use cases:

COPY / BRAND WRITING
No filler phrases. No exclamation marks unless explicitly requested. Never use the words: "transform", "elevate", "curate", "journey", "innovative", "cutting-edge", "seamless", "robust", "leverage", "synergy". No introductory sentences explaining what you're about to do. No summarizing what you just wrote at the end. Do not start sentences with "I". No bullet points unless specifically asked.
FACTUAL / RESEARCH WRITING
No hallucinated statistics or citations. Do not fabricate sources. No hedging phrases like "it's worth noting that" or "it's important to consider". No repetition of information already stated. No vague qualifiers. No padding. Cite only what you can confirm. If uncertain, say so directly.

Questions people ask

Does SDXL need a negative prompt?

No. In Diffusers it is marked optional, and the library’s own SDXL text-to-image example runs with a prompt alone (sources in the Stable Diffusion section above). Add one when you see a problem you want gone.

What is the best negative prompt for SDXL?

There is no single best one. Start with the short general-quality line above, add a targeted line only for problems you can see, and keep anything that clashes with your prompt out of it.

Do negative prompts work on SDXL Turbo or FLUX?

Not on SDXL Turbo at its recommended setting: Diffusers says to set guidance to 0.0 for it, and the negative prompt is ignored without guidance. FLUX.2 has no negative prompt: Black Forest Labs says “FLUX.2 does not support negative prompts.”24

What is EasyNegative?

A negative embedding: Diffusers calls it “a widely used negative embedding that contains multiple learned negative concepts.” You load the file and type its trigger word in the negative prompt (source in the Stable Diffusion section above).

SEE MORE: Negative prompts by problem sorts the same kind of lists by what went wrong, and no negative prompt field? covers how to keep things out of a picture where there is none.

QUICK CHECK · 3 QUESTIONS

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

1. According to the page, what should you check before pasting a negative prompt?

  • Whether your model actually reads a negative prompt
  • Whether the list is long enough to cover every possible flaw
  • Whether it includes quality words like masterpiece
Show the answer

Answer: Whether your model actually reads a negative prompt. The page says models disagree on whether a negative prompt exists at all, so check your model before pasting anything.

2. What does the page say Runway warns can happen if you include a negative prompt?

  • It is quietly ignored with no effect
  • It is counted as an extra generation
  • It may produce the opposite of what you asked
Show the answer

Answer: It may produce the opposite of what you asked. The page quotes Runway’s guide: negative prompts are not supported, and including one may result in the opposite happening.

3. On a model with no negative field, such as FLUX, what does the page suggest you do instead?

  • Paste the negative list at the very end anyway
  • Describe what you do want, such as visible pores instead of no plastic skin
  • Turn the guidance setting up as high as it goes
Show the answer

Answer: Describe what you do want, such as visible pores instead of no plastic skin. The page says the vendors’ own advice is to rewrite each exclusion as a description of what you want, placed early in the prompt.

The starter lists, the shopping-trolley comparison and the plain-words description of guidance in the Stable Diffusion section are this site’s own ways of explaining, not claims from these sources. InvokeAI’s support pages could not be reached on 12 Oct 2026, so InvokeAI is not covered here.

SOURCES
  1. 1Black Forest Labs, FLUX.2 [pro] & [max] prompting guide, re-read at source 10 Sep 2026. BFL’s reason“even when they can process them, AI models generally struggle with negation” ↑
  2. 2FLUX.1 [dev] model card · FLUX.1 [schnell] model card · FLUX.2 prompting guide; all read at source 16 Sep 2026.↑
  3. 3Google Cloud, Generate videos from text (Gemini Enterprise Agent Platform), last updated 10 Oct 2026, read at source 12 Oct 2026“NEGATIVE_PROMPT: Optional: A string value that describes content that you want to prevent the model from generating.” · Google, Generate videos with Veo (Gemini API), read at source 12 Oct 2026: no negative-prompt parameter in its table. ↑
  4. 4Google, Gemini Omni (Gemini app overview), undated, read at source 13 Oct 2026“Gemini Omni is our latest video editing and generation model that will replace Veo in the Gemini app.” ↑
  5. 5Google, Gemini API deprecations, last updated 12 Oct 2026, read at source 13 Oct 2026veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview, shutdown “October 22, 2026”, recommended replacement “gemini-omni-1.1-flash”; the page notes that its dates “indicate the earliest possible dates on which a model might be retired”. ↑
  6. 6Google, Video generation in the Gemini API, read at source 3 Nov 2026“Veo 3.1 Preview (veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview) in the Gemini API was shut down on October 22, 2026; migrate to Gemini Omni Flash (gemini-omni-1.1-flash) or the GA models available through the Gemini Enterprise Agent Platform to avoid service interruptions.” ↑
  7. 7Google Cloud, Ultimate prompting guide for Veo 3.1, 16 Oct 2025, re-read at source 12 Oct 2026.↑
  8. 8Google, Gemini API image generation docs, last updated 9 Oct 2026, read at source 12 Oct 2026 (no negative-prompt parameter on the page) · Google Developers Blog, How to prompt Gemini 2.5 Flash Image Generation for the best results, 28 Aug 2025, read at source 12 Oct 2026“Instead of saying ‘no cars,’ describe the desired scene positively”. ↑
  9. 9Google, Gemini API image generation, read at source 12 Oct 2026, multi-turn editing example.↑
  10. 10Dreamina (ByteDance), Seedance 2.5 Prompt Guide, 18 Aug 2026 — the vendor’s own platform: “Keep them short, concrete, and grouped at the end. Do not write a long generic blacklist that competes with the scene description.” Luma AI, Seedance 2.5 Complete Guide, 30 Jul 2026: “avoid giant generic negative lists”. Melies, Seedance 2.5 Prompt Guide, 2026: “A giant generic negative prompt often contains conflicts.” First checked 25 Aug 2026; all three re-read at source 23 Sep 2026. The apparent disagreement resolves: ByteDance does say to write the wanted and unwanted result together — their template has slots for both, Preserve then Avoid. What they do not endorse is a long generic exclusion list, which is what Luma and Melies warn against. Both are saying the same thing: a short, concrete Avoid clause at the end, not a dump.↑
  11. 11Runway, Gen-4 Image Prompting Guide and Gen-4 Video Prompting Guide, read in a browser 17 Sep 2026. An earlier version of this paragraph gave Runway’s structure as subject → action → setting → camera → motion over time → style; the guides do not give that order.↑
  12. 12AUTOMATIC1111, stable-diffusion-webui wiki: Features, last edited 15 Dec 2024, read at source 12 Oct 2026“This works by using the negative prompt for unconditional conditioning in the sampling process instead of an empty string.” ↑a“Allows you to use another prompt of things the model should avoid when generating the picture.” ↑bthe Extra networks table gives Textual Inversion’s prompt syntax as “embedding’s filename”, with its cards in the embeddings directory. ↑c
  13. 13Hugging Face, Diffusers: Text-to-image, read at source 12 Oct 2026“Just like how a prompt guides generation, a negative prompt steers the model away from things you don’t want the model to generate. This is commonly used to improve overall image quality by removing poor or bad image features such as ‘low resolution’ or ‘bad details’.” ↑
  14. 14Hugging Face, Diffusers: StableDiffusionPipeline, read at source 12 Oct 2026guidance_scale “defaults to 7.5”; “Guidance scale is enabled when guidance_scale > 1.” negative_prompt: “Ignored when not using guidance (guidance_scale < 1).” ↑
  15. 15Hugging Face, Diffusers: Stable Diffusion XL, read at source 12 Oct 2026“SDXL combines a second text encoder (OpenCLIP ViT-bigG/14) with the original text encoder to significantly increase the number of parameters”. ↑a“Diffusers also lets you specify negative conditions about an image’s size to steer generation away from certain image resolutions”. ↑b
  16. 16Hugging Face, Diffusers: StableDiffusionXLPipeline, read at source 12 Oct 2026negative_prompt_2 “The prompt or prompts not to guide the image generation to be sent to tokenizer_2 and text_encoder_2. If not defined, negative_prompt is used in both text-encoders”. guidance_scale “defaults to 5.0”. ↑
  17. 17Hugging Face, Diffusers: Stable Diffusion XL Turbo, read at source 12 Oct 2026“Make sure to set guidance_scale to 0.0 to disable, as the model was trained without it.” ↑
  18. 18Civitai, Prompt enhancement (developer docs), read at source 12 Oct 2026“Most useful on SD1/SDXL where negative prompts carry weight; often unnecessary on Flux/video.” ↑
  19. 19ComfyUI, Text to Image Workflow, read at source 12 Oct 2026KSampler parameter “negative”: “Negative prompt condition encoding.” “Suppresses unwanted content.” ↑
  20. 20Stability AI, Developer Platform API reference, read in a browser 12 Oct 2026negative_prompt “string (NegativePrompt) <= 10000 characters” “A blurb of text describing what you do not wish to see in the output image. This is an advanced feature.” ↑
  21. 21Civitai Education, Civitai’s Prompt-Crafting Guide: Part 1 – Basics, last updated 13 Mar 2025, read at source 12 Oct 2026“The solution is to utilize the dedicated Negative Prompt section, to remove undesired features from our images.” ↑a“low res, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, ugly”. ↑b“Repeating the same keywords does not necessarily improve results and may lead to distortions.” “Very high weight values, or many weighted values, can cause severe artifacting and incoherent outputs.” ↑c“can confuse the model and produce inconsistent results.” ↑d
  22. 22Ban, Wang, Zhou, Cheng, Gong and Hsieh, Understanding the Impact of Negative Prompts: When and How Do They Take Effect?, arXiv, 5 Jun 2024, read at source 12 Oct 2026“Delayed Effect: The impact of negative prompts is observed after positive prompts render corresponding content.” “Deletion Through Neutralization: Negative prompts delete concepts from the generated image through a mutual cancellation effect in latent space”. ↑
  23. 23Hugging Face, Diffusers: Textual Inversion, read at source 12 Oct 2026“Textual Inversion can also be trained to learn negative embeddings to steer generation away from unwanted characteristics such as ‘blurry’ or ‘ugly’. It is useful for improving image quality.” “EasyNegative is a widely used negative embedding that contains multiple learned negative concepts.” “Pass the token to negative_prompt in your pipeline to activate it.” ↑
  24. 24Black Forest Labs, FLUX.2 prompting guide, read at source 12 Oct 2026“FLUX.2 does not support negative prompts.” ↑

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