Negative prompts explained: how each model handles them, and which ones actually help

A negative prompt is a list of things you want the model to steer away from. It sounds simple, but it works very differently from one generator to the next, and in some tools it doesn’t exist at all. Knowing which situation you are in saves a lot of copying and pasting of long “bad quality” lists that do little or nothing.

Why you can’t just write “no cars”

Image models read prompts as a bag of associations, not as instructions. When you write “a quiet street with no cars”, the word cars is still in the text, and the model is still being nudged towards things that look like cars. The word no carries very little weight against a concrete noun.

That is the whole reason negative prompts exist: they give you a separate place to name the unwanted thing, so it pushes the image away instead of towards.

How Stable Diffusion uses a negative prompt

Stable Diffusion generates images with a technique called classifier-free guidance. At every step, the model makes two predictions: one conditioned on your prompt and one conditioned on an “empty” prompt. It then moves the image further in the direction of your prompt and away from the empty one. The CFG scale controls how hard it pushes.

A negative prompt simply replaces that empty prompt. Instead of moving away from “nothing in particular”, the image moves away from whatever you typed in the negative field. Two practical consequences follow:

• It costs nothing extra. The second prediction happens anyway, so a negative prompt doesn’t slow generation down.
• It scales with CFG. At a very low CFG scale, the negative prompt has little effect. At a high scale, it has a strong one, and an aggressive negative list can flatten colours or strip out detail along with the thing you wanted gone.

Weights work in the negative field too, so (watermark:1.3) pushes harder against watermarks than a plain watermark. Our guide to Stable Diffusion prompt weights covers that syntax in detail, and the Stable Diffusion prompt generator writes a positive and a negative prompt for you from an uploaded image.

How Midjourney handles --no

Midjourney has no separate negative field. Instead, it has the --no parameter, which goes at the end of the prompt with the other parameters:

a quiet cobbled street at dawn, shuttered shops, soft mist --no cars, people --ar 3:2

Midjourney documents --no as the equivalent of giving that term a negative weight in a multi-prompt, so --no cars behaves like cars::-.5. It works well for concrete things: objects, colours, text, frames. It is less reliable for abstract qualities such as “ugly” or “bad composition”, because there is no clear visual concept to push against.

Keep the list short. Each extra term dilutes the others, and listing ten unwanted things often gives worse results than fixing the positive description. The Midjourney parameters guide explains how --no interacts with the rest of the parameter block.

Flux, DALL-E and Gemini: describe what you want instead

The distilled Flux models most people use (such as Flux.1 dev and schnell) were trained to work without classic classifier-free guidance, so most Flux interfaces don’t offer a negative prompt field. Some local workflows add one back with extra guidance tricks, at the cost of slower generation, but it isn’t the normal way to use Flux.

DALL-E and Gemini (including Nano Banana) also take a single natural-language prompt. You can write “no text in the image” and these models often respect it, because they are better at following sentences than older diffusion models. Still, the most reliable approach is to describe the result positively:

• Instead of: a beach with no people
• Write: an empty beach at sunrise, untouched sand, a single line of footprints washed away by the tide

The positive version gives the model something to draw, rather than something to avoid. That habit carries over to every tool; the Flux prompt generator and Nano Banana prompt generator both write prompts this way.

Which negatives actually help

Negative prompts work best when they name something visual and specific that the model genuinely tends to add. Useful categories include:

Unwanted objects and overlays

text, watermark, signature, logo, caption, frame, border. Models learned from images that often contained these, so they add them unprompted. Naming them in the negative field is one of the clearest wins.

Style drift

If you want a photograph and keep getting something that looks painted, add painting, illustration, 3d render, cartoon to the negative. If you want an illustration and it keeps going photographic, flip it: photograph, photorealistic.

Specific recurring problems

When a particular image keeps producing a problem you can see, name that problem: lens flare, heavy vignette, oversaturated, blurry background (when you want everything sharp), duplicate subject. These are targeted, and you can check whether they worked.

Quality tags on models trained with them

Terms like worst quality, low quality do have an effect on some community fine-tunes, particularly anime-style models whose training captions included those exact quality tags. On models trained on ordinary natural-language captions, the same words have much less to latch onto.

Which negatives are mostly placebo

Long negative lists get copied from prompt to prompt for years, and many entries do far less than people assume:

• Anatomy wish lists. bad anatomy, extra fingers, mutated hands, deformed rarely fixes hands. Hand quality depends mostly on the model, the resolution and the pose. Naming fingers can even keep the concept of hands active in the guidance.
• Abstract judgements. ugly, bad art, boring, amateur don’t correspond to a consistent visual pattern the model can move away from.
• Contradictions with your own prompt. A negative of blurry on a prompt that asks for shallow depth of field works against the soft background you asked for.
• Huge lists. A negative prompt is processed like a positive one. Fifty terms compete for attention, so the few that matter get diluted.

A good test: if you can’t picture what a negative term looks like, the model probably can’t either.

Worked examples

Product shot in Stable Diffusion

Positive: a ceramic coffee mug on a pale oak table, soft window light from the left,
minimal studio product photo, sharp focus, neutral beige background

Negative: text, logo, watermark, cluttered background, illustration, 3d render

Every negative term names something a product shot tends to pick up. There is no anatomy list, because there are no people.

Landscape in Midjourney

an alpine lake at first light, mirror-still water, pine forest along the shore,
pale pink sky, fine-art landscape photograph --no boats, people, text --ar 16:9 --style raw

Three concrete exclusions, all of them things that commonly appear in lake scenes.

The same idea in Flux

An alpine lake at first light. The water is perfectly still and reflects a pale pink sky.
Pine forest lines the shore, and the scene is completely empty: no boats, no people,
no buildings. Fine-art landscape photograph, natural colours.

Flux has no negative field, so the exclusions become part of the description, framed around what the scene is: empty and still.

A quick checklist

1. Check whether your tool has a negative field (Stable Diffusion), a parameter (--no in Midjourney), or neither (Flux, DALL-E, Gemini).
2. Fix the positive prompt first. Most unwanted elements disappear once the description is specific.
3. Name concrete, visual things: objects, overlays, styles, specific artefacts you have actually seen.
4. Keep the list short, roughly a handful of terms, and remove anything that contradicts your positive prompt.
5. Drop abstract quality words unless your specific model was trained on them.
6. Change one negative term at a time so you can see which one did the work.

If you’re starting from a reference image, the image to prompt tool gives you a clean positive description to build on, which is usually the bigger win.