Stable Diffusion Negative Prompts Explained (+ Copy-Ready Lists)

You wrote a careful prompt, hit generate, and SDXL delivered exactly what you feared: a portrait with six fingers, a product shot with gibberish text stamped across it, skin so smooth it looks injection-molded. Rerolling won’t save you — the model isn’t failing at random, it’s drifting toward defaults you never told it to avoid. The negative prompt is the tool built for exactly this. It’s the control Midjourney users don’t have (their --no parameter is a pale imitation) and the one Stable Diffusion users most consistently underuse, pasting the same fifty-term block into every generation or skipping it entirely. This guide explains what a negative prompt actually does, gives you copy-ready lists for the five most common use cases, and — more valuable than any list — teaches the habit of turning what you see going wrong into the exact words that stop it.

The clearest statistic in our 200-prompt library: 100% of the 48 tested SDXL entries required a negative prompt to reach publishable quality, on a median recipe of 30 steps at CFG 6.5 (full corpus data).

What a negative prompt actually does

A negative prompt is a second, full prompt that runs alongside your main one. At every denoising step, Stable Diffusion predicts the image twice — once conditioned on your positive prompt, once on your negative — and then steers the result away from the negative prediction and toward the positive one. “Blurry” in the negative field isn’t a flag or a filter; the model genuinely imagines a blurry version of your scene and moves in the opposite direction.

Two consequences follow. First, negatives work best on things the model can visualize: “extra fingers” is a concrete concept it can steer away from, while “bad” is not. Second, the strength of that steering is governed by your CFG scale. Here’s the honest version: CFG multiplies the gap between the negative and positive predictions, so at CFG 7 your negatives have real pull, at CFG 3 they barely register, and at CFG 12 both prompts are enforced so hard that images turn crunchy and oversaturated. If your negatives seem to do nothing, check CFG before adding more words.

What a negative prompt is not is a magic quality switch. Stuffing “worst quality, low quality” into the box doesn’t unlock a hidden good-image mode — it just nudges sampling away from artifacts, and only the artifacts you name.

The universal baseline

If you only ever use one negative prompt, use this:

text, watermark, logo, signature, blurry, lowres, jpeg artifacts, deformed, bad anatomy, extra fingers, extra limbs, mutated hands, disfigured, worst quality

This covers the failure modes that show up across nearly every subject: stray text and watermarks the model learned from stock photos, compression fuzz, and the anatomy errors that plague any image containing hands. For quick experiments and non-human subjects, this baseline is genuinely enough — add to it only when a render shows you a problem it doesn’t already cover.

Copy-ready lists by use case

Different subjects fail differently, so a good negative prompt is use-case-shaped. Start from the block that matches your project.

Photorealistic portraits:

plastic skin, waxy skin, airbrushed, doll-like face, cross-eyed, asymmetrical eyes, bad teeth, extra fingers, fused fingers, 3d render, cgi, cartoon, anime, overexposed, harsh flash

Portraits fail at skin and eyes: the “plastic, waxy, airbrushed, doll-like” cluster fights SDXL’s tendency to over-smooth faces into mannequins, while “cartoon, anime, 3d render” keeps the output pinned to photography instead of drifting toward illustrated training data.

Anime and cel-shaded art:

photorealistic, photo, realistic skin texture, western cartoon, 3d render, blurry lineart, muddy shading, off-model face, sketch lines, extra digits, bad hands, artist signature

This is nearly the mirror image of the portrait list — here realism is the contaminant. “Blurry lineart” and “muddy shading” protect the two things that make cel art read as cel art, and “western cartoon” stops the style sliding toward a different animation tradition entirely.

Product shots:

cluttered background, reflections of people, fingerprints, dust, scratches, warped edges, distorted proportions, gibberish text, fake logo, harsh shadows, motion blur

Product photography lives or dies on cleanliness: “warped edges, distorted proportions” guard the geometry that makes an object look manufactured rather than melted, and “gibberish text, fake logo” suppress the fake branding Stable Diffusion loves to invent on packaging.

Painterly and watercolor:

photorealistic, 3d render, hard outlines, digital airbrush, smooth gradient, flat color fill, oversaturated, muddy colors, lens flare

Traditional-media prompts collapse when digital tells creep in — “hard outlines, digital airbrush, smooth gradient” are precisely the marks a real brush doesn’t leave, and “muddy colors” prevents the overworked-palette look that ruins watercolor washes.

Landscapes and wallpapers:

people, buildings, power lines, vehicles, tilted horizon, fisheye distortion, oversharpened, hdr halos, banding, vignette, low detail foreground

For pure nature scenes, the biggest offenders are human intrusions — “people, buildings, power lines” — plus the processing artifacts that betray a fake photo: “hdr halos” and “banding” matter especially at wallpaper sizes, where smooth skies make both painfully visible.

Reading failures and writing negatives

The lists above are starting points. The skill that actually compounds is diagnostic: generate, look at what went wrong, name the artifact you see, and add it verbatim to the negative prompt. Not “make it better” — the literal thing. Doubled wings? Add “extra wings”. A second head sprouting from a coil? Add “extra heads”. The entries in our Stable Diffusion library carry their tested negatives for exactly this reason — here’s the habit at work in three of them.

A stormy sea full of serpent coils gives SDXL many chances to grow anatomy where a coil meets the foam:

Why it works: “extra heads” is in this negative because coils breaking multiple wave crests frequently sprout a second head at the waterline — a failure specific to this composition — while “photorealistic” holds the image to its dark oil-painting palette.

A hummingbird mid-hover is a wing-doubling machine:

Why it works: “extra wings” targets this prompt’s single most common failure — SDXL doubling the blurred wings mid-hover — and “hard outlines, muddy colors” defend the wet-on-wet washes that make it read as watercolor at all.

And a character portrait puts everything on the hands and face:

Why it works: “deformed hands, extra fingers” back up the positive prompt’s “helmet under one arm” (busy hands are countable hands), and “photorealistic skin” keeps the face inside the painterly, visible-brushstroke style instead of snapping to photo-real.

Notice the pattern: each negative answers a failure this specific prompt produces. That’s the standard to hold your own negatives to.

Mistakes that make negative prompts worse

1. Negative-stuffing. Fifty terms don’t give you fifty protections — they dilute each other, because the steering budget is shared across everything the negative prompt describes. Ten targeted terms beat fifty generic ones.
2. Negating what you affirm. “Detailed” in the positive and “highly detailed” in the negative forces the model to split the difference. Audit for contradictions before blaming the sampler.
3. Copying random mega-lists. Those blocks circulating on forums were tuned for specific checkpoints, and terms like “anime” in a copied list will quietly fight you when anime is what you want.
4. Expecting negatives to rescue a weak positive prompt. A negative prompt removes; it cannot add. If the image is missing subject, style, or lighting, fix the positive prompt first — negatives polish a good render, they don’t create one.

Do negative prompts work on every model?

No. Stable Diffusion and SDXL support them natively — a true second prompt with its own text field in every major UI. Midjourney has no negative field; the --no parameter appended to the prompt is the closest equivalent. Flux.1-dev and DALL·E 3 have no true negative channel at all: writing “no text” in those prompts often increases the odds of text appearing, so phrase positively instead — “a clean, unlabeled bottle” rather than “no logo”.

Where to go next

Grab a tested starting point from the Stable Diffusion prompt library — every entry ships with its negative prompt and the settings it was verified at. If cel-shaded work is your target, the anime style hub collects prompts where the anti-realism negatives above earn their keep, and our best anime AI prompts post shows them applied across a full set. Copy a baseline, generate, read the failure, name it — that loop is the whole skill.