A negative prompt tells Stable Diffusion what to steer away from. You list the failure modes you keep seeing — extra fingers, watermarks, mushy backgrounds — and the model pushes the image away from them. Used well, negative prompts in Stable Diffusion fix more problems than another round of positive descriptors ever will. Used badly, they strip the life out of your image or do nothing at all.

This page explains how the mechanism actually works, then hands you ready-to-paste lists: a sensible general baseline, plus targeted sets for portraits, anatomy, and text artifacts. Copy what fits, drop the rest.

What a negative prompt actually does #

Under the hood, classifier-free guidance runs your render twice on each step: once toward the positive prompt, once toward the negative. The model then moves in the direction that increases the positive and decreases the negative. So the negative prompt is not a blocklist or a filter sitting on top of the image. It is a second prompt, pulling in the opposite direction with the same force the positive pushes.

Two things follow. First, weight matters here exactly like it does in the positive prompt; (blurry:1.4) pushes away from blur harder than a bare blurry. Second, and people learn this the hard way, an overstuffed negative prompt removes things you wanted. List shadows and you flatten your lighting. List dark and your moody night scene turns to daylight. The negative prompt is a scalpel, not a fire hose.

If your images suddenly look washed out, pale, and lifeless, your negative prompt is too aggressive. Cut it in half. Most of the time three or four well-chosen terms beat a thirty-token wall.

SDXL changed the math

The long negative-prompt lists you see in old SD 1.5 guides were partly compensating for a weaker model. SDXL and newer fine-tunes need far less. On SDXL, start minimal — often just quality and artifact terms — and add a negative token only when you can name the specific problem in front of you. Pasting a 1.5-era mega-list into SDXL frequently makes things worse, not better.

General baseline (start here) #

This is a safe, model-agnostic starting point. It targets the universal failure modes without touching color, mood, or lighting.

lowres, worst quality, low quality, jpeg artifacts,
blurry, out of focus, deformed, disfigured,
watermark, signature, text, username,
cropped, out of frame

That is enough for most work. Resist the urge to bolt on more until a specific problem shows up. When one does, reach for the targeted lists below.

Portrait and face negatives #

Faces are where the model fails most visibly, and where viewers are least forgiving. This set cleans up the usual portrait gremlins: asymmetric eyes, plastic skin, weird teeth.

(deformed face:1.2), asymmetric eyes, cross-eyed,
mutated, ugly, plastic skin, waxy skin,
extra teeth, bad teeth, long neck,
makeup smear, airbrushed, doll-like,
disproportionate, bad proportions

Two notes. airbrushed and plastic skin are the ones that buy you realistic skin texture — they push the model off the over-smoothed look it defaults to. And if your subject keeps coming out younger than intended, add child, childish to the negative; if older, the reverse. Tune to the failure you actually see.

Anatomy and hands #

Hands are the model’s oldest enemy. SDXL improved them, but they still drift. This is the targeted set for limbs, fingers, and overall body coherence.

(extra fingers:1.3), (fused fingers:1.2), missing fingers,
too many fingers, mutated hands, malformed hands,
extra limbs, missing limbs, extra arms, extra legs,
disconnected limbs, floating limbs,
bad anatomy, bad proportions, twisted body

Honest expectation: negatives reduce hand failures, they do not eliminate them. The reliable fix is workflow, not wishing. Generate a batch, pick the frame with the best hands, then inpaint or run an ADetailer / hand-detail pass on that one image. A clean negative prompt raises your odds per generation; targeted inpainting closes the gap.

Why weighting the anatomy terms helps

Notice the (extra fingers:1.3) weighting above. Finger problems are stubborn because the model genuinely struggles to count, so a little extra push earns its keep here even though you would not weight, say, watermark. Keep it to 1.2–1.3. Crank it to 1.8 and the hands sometimes vanish into stumps, which is its own kind of broken.

Text, logos, and watermark artifacts #

Stable Diffusion loves to scrawl fake text, garbled signatures, and stock-photo watermarks across an image, especially on photographic styles. This set scrubs them.

text, words, letters, caption, title,
watermark, signature, stamp, logo, label,
copyright, artist name, username, web address,
gibberish text, error, ui, frame border

If a stubborn watermark survives, weight it: (watermark:1.3), (text:1.3). And if you are generating something that legitimately needs text — a poster, a sign — do not put text in the negative at all, or you will sabotage yourself. Match the negative to the job.

Embeddings: negatives as a single token #

If you find yourself pasting the same long negative prompt into every render, there is a tidier option. Negative textual-inversion embeddings bundle a whole set of “bad image” concepts into one keyword you drop into the negative field. You will see community ones with names that signal their purpose; you trigger them by typing the embedding’s filename as a single token. One word in the negative does the work of a dozen.

They are convenient, but treat them like any other negative term: with restraint. A strong embedding can over-correct just like an overstuffed list, draining color or homogenizing faces toward one look. Add it, render, and check that the image still has the contrast and character you wanted. If it flattens things, lower its weight — (embedding-name:0.8) — or drop it for a hand-written list on that particular piece. The goal is a cleaner image, not a token you paste on autopilot.

A reasonable default for many users: a short hand-written baseline for general work, and one trusted embedding kept in reserve for portrait and full-body renders where anatomy and skin tend to fail. That covers the common cases without turning every prompt into a guessing game about which of thirty tokens is dulling your output.

How to actually use these #

  1. Start with the general baseline. Render. See what breaks.
  2. Add one targeted set that matches the failure — the portrait list for a face shot, the anatomy list for full-body.
  3. Weight only the stubborn terms, 1.2 to 1.4, never higher.
  4. If the image looks washed out, you went too far. Remove tokens until color and contrast return.
  5. Save your working negatives as presets. A reusable portrait negative and a reusable anatomy negative cover most of what you make.

Negative prompts are the cheapest quality upgrade in Stable Diffusion, but only when they stay sharp and specific. Lead with the positive prompt, keep the negative lean, weight the few terms that fight back, and inpaint the rest. Pair these lists with a well-structured positive prompt and your reject rate drops fast. The prompt generator at ArtPrompts Generator pairs a clean positive skeleton with a matching negative so you are not assembling both from scratch every time.