Negative prompts: how to use them well, tool by tool

You've surely come across this list, copied from tutorial to tutorial: ugly, deformed, extra fingers, mutated hands, poorly drawn face, bad anatomy, blurry, low quality, worst quality, watermark, signature..., running three lines. A good part of it dates back to the first Stable Diffusion models, and pasting it everywhere is like taking every medicine in the cabinet and hoping one of them works. You won't know what helped, and some terms can pull your image in a direction you didn't choose.

A negative prompt works better as a correction tool: you generate, you spot one specific flaw, you exclude it.

What a negative prompt really does

On Stable Diffusion, this field is not a filter that erases objects after the fact. According to the AUTOMATIC1111 documentation, at each step the sampler compares what your positive prompt produces with what your negative prompt produces, then pushes the image toward the first and away from the second. In practice, that changes several things.

The negative influences the whole image, not just the targeted object. Removing people from a street can also alter the setting. It then depends on the strength of the guidance: when the CFG is very low, it loses its effect, and the Diffusers documentation for SD3 states that it is ignored without guidance. Finally, it guarantees nothing: an excluded element can still show up, especially if your positive prompt calls for it indirectly. Asking for a beach in midsummer with people in the negative is still hard to pull off.

What changes depending on the tool

There is no universal syntax. Many disappointments come from a method applied to the wrong tool.

Three families of terms that work

Write your negative with the names of concrete flaws or objects. Three families cover the vast majority of needs.

Stray elements. Anything that crashes the image uninvited: text, watermark, signature, logo. This is the most profitable use, because these elements often come from the training images.

The wrong type of rendering. If you want a photo and the image drifts toward illustration, write illustration, 3d render, cartoon. In the other direction, photo, photorealistic helps keep a drawing flat.

The flaw you spotted in your output. Plastic-looking plate, oversaturated sky, background too busy: write what you see, for example plastic texture or oversaturated sky.

Here's a starting point for a real estate agency visual, on SDXL. Paste each part in its own field:

Positive:
photograph of [ROOM] in a [TYPE OF HOME], natural daylight from a large window, neutral tones, tidy space, wide-angle view at eye level

Negative:
text, watermark, logo, oversaturated colors, fisheye distortion, people

Important: a generated image doesn't show the actual property. In a listing, it can only be used to illustrate a potential, clearly labeled as such, and never to represent the home that is for sale.

Rephrasing in the positive: often the best option

When possible, describe the state you want instead. The model has far more to grab onto with a scene than with an absence.

On Midjourney, writing the negation into the sentence sets a trap: a bakery window without cakes can bring in cakes, since the word is still in the prompt. On FLUX, rephrasing is the only solution. Take the example of a florist who wants a bouquet on a plain background for her social media.

Weak prompt: bouquet of tulips, no vase, no background, no shadows, not cluttered

Corrected prompt:

studio photograph of a hand-tied bouquet of [FLOWERS], wrapped in [MATERIAL], lying flat on a plain [COLOR] paper background, soft even light from above, minimal soft shadow, overhead view, centered composition, [2-COLOR PALETTE] tones --no vase, text, logo

Everything that was an absence has become a description: bouquet lying flat, plain background, even light. The final --no keeps only what really matters. Remove that parameter if you're not using Midjourney.

Mistakes to fix right away

The negative that contradicts the positive. sharp details on one side and detailed in the negative on the other: the two cancel each other out. Reread both fields together.

Excessive weight. On Stable Diffusion, (worst quality:2) can distort the result and stiffen the image. Stay close to 1, and go up to 1.2 or 1.3 only if a flaw persists.

A negative inherited from another model. A list made for SD 1.5 has no reason to suit SDXL or SD 3.5. Some terms like worst quality were popularized by specific community models and may have no effect elsewhere. For choosing a model and settings, the Stable Diffusion guide helps you figure out where you stand.

Negative embeddings deserve a mention: these small files bundle common flaws into a single word. They can be effective, but they're just as opaque as the big lists, and they depend on the model family. Test them like everything else.

The ten-minute testing method

Don't judge on a single image, because chance could fool you. With a fixed seed on Stable Diffusion, proceed like this.

1. Generate your image with no negative at all.
2. Note the most bothersome flaw, just one.
3. Add one or two terms that target it, and rerun with the same seed.
4. Look at what changed, including what you hadn't asked to change.
5. Repeat with a different seed to check that the effect holds.

On Midjourney, --seed helps too, bearing in mind that an identical result isn't guaranteed. The spirit is the same as when you test a prompt for text: change one thing, compare, keep what works.

You'll come out of this with a negative of five to ten words, each with a reason to be there. Take your latest failed image, spot its main flaw, and start with that single word.