How to Write Negative Prompts: A Common Checklist and the Reasoning Behind It

Let's define this clearly first: a negative prompt isn't "the opposite-word version of a prompt" — it's telling the model "push the probability weight for these directions down a bit." Roughly speaking, it adds an exclusion boundary to the sampling process — it's not deletion, it's down-weighting.

Understanding this distinction matters, because it determines how a negative prompt should be written and how much of it is actually useful.

When you need one

It mostly depends on which category of model you're using. Traditional diffusion models (things like Stable Diffusion, which have a dedicated negative-prompt input box in the interface) rely heavily on negative prompts, because they don't have any natural language understanding of "what not to include" — the whole mechanism relies on this field to forcibly suppress probability.

Instruction-following models (like GPT-based image generation, or Nano Banana-style models that lean on natural language understanding) work differently. They're closer to "tell it clearly what you want and it does it" — negative prompts have limited effect on this type of model, and sometimes writing too much negative description gets interpreted by the model as a hint that "these elements should appear in the image," backfiring entirely. Roughly speaking — the more you emphasize "no red," the more it tends to lean toward red.

So the rule of thumb is simple: does the interface have a dedicated negative-prompt input box? If yes, follow the old rules. If not, fold your requirements into the regular description instead — say directly "clean background, no clutter," rather than negating it separately.

A common negative-word checklist

This list is organized by scenario — don't copy the whole thing wholesale, just pull what you need:

• Portrait general: extra fingers, deformed limbs, blurry facial features, vacant expression, over-smoothed skin
• Image quality: low resolution, noise, compression artifacts, overexposure, underexposure
• Composition: bad cropping, subject off-center, extra borders, watermark, text
• Style contamination: cartoonish (when you're going for realism), greasy filter look, over-saturation

Comparing these, a pattern emerges: negative prompts mostly solve for "the model's systematic biases," not aesthetic judgment. Broken fingers, extra limbs — these are old, recurring flaws of generative models, and this kind of term works reliably in a negative prompt. But subjective words like "ugly" or "not pretty" are basically useless in there — the model has no way to translate "ugly" into a concrete direction for pixel adjustment.

On the hand-distortion issue specifically, there's a separate piece — A Practical Fix for Hand Distortion — that's more directly on-target than a negative prompt.

When you don't need one

Three situations where you can skip negative prompts entirely:

• You're using a newer model that leans on natural language understanding — stating your requirements positively is enough
• Generating small, simply-composed images like avatars — the error rate is already low to begin with, so a negative prompt isn't worth the cost
• You're already doing image-to-image with a reference image — the composition is already locked down by the reference, so there's not much room left for a negative prompt to correct anyway

Roughly speaking, it's an engineering judgment call: the payoff from a negative prompt scales with the freedom in the generation task. The more freedom (pure text-to-image, no reference, complex scenes), the more the negative prompt matters; the more that freedom is already locked down by a reference image or a detailed description, the more the negative prompt's marginal effect approaches zero.

A real comparison

For the same portrait prompt, the difference between adding a negative prompt or not shows up most clearly in complex-pose scenarios. Take a task like "group photo, multiple people" — without a negative prompt, the odds of overlapping limbs going wrong are high; add "extra limbs, fused fingers, overlapping faces" and the error rate drops noticeably. But for a single person facing the camera, half-body shot, adding or skipping the negative prompt barely shows a visible difference — the error rate was already low to begin with.

The conclusion is one line: negative prompts are for countering a model's known systematic weaknesses, not for expressing your aesthetic preferences. Get that boundary clear, and you'll stop piling on random words for no reason. To try it directly, the homepage has a prompt tool entry point.