How to Write Negative Prompts: Common Problems, Word Lists, and Their Limits

A negative prompt is a string of English words placed in the negative/exclude box, telling the model what should ideally not appear. Its mechanism is usually explained this way: during generation the positive and negative prompts guide the image together, and at every denoising step the model nudges the picture away from whatever the negative side describes; this relies on the CFG guidance scale, and with the scale at zero negative words barely act. Negative words are therefore a correction signal, not a command — they cannot guarantee an element disappears completely, only noticeably lower the chance of it showing up.

Which models respond matters first. The Stable Diffusion family (SD 1.5, SDXL and their many derivatives) is the most sensitive to negative words and usually shows the clearest effect; most local packages and online SD platforms provide a dedicated negative box, named slightly differently across platforms — negative prompt or exclude — but the idea is the same. Midjourney has no separate negative box — you express it with the --no parameter, for example adding --no text, watermark at the end of your prompt. The Flux family generally does not need negative prompts: it uses distilled guidance and its official interfaces usually have no negative input at all, so forcing words in helps little and may even cause conflicts. Conversational image models such as GPT-4o likewise have no negative field; write exclusions directly into the positive prompt.

The most effective way to write negatives is to patch wherever the problem appears, not to recite one giant list. Deformed hands are the most common failure, because hand poses vary endlessly and hands in training images are often holding or hiding something; the usual additions are: extra fingers, missing fingers, fused fingers, mutated hands, bad hands, deformed hands. Broken faces often show up in distant or multi-person compositions — a small face occupies few pixels, so its details break down more easily; general additions are: bad anatomy, deformed face, disfigured, cross-eyed, poorly drawn face.

For quality issues, the usual words are: blurry, lowres, low quality, jpeg artifacts, oversaturated. For structural errors such as extra limbs, general additions are: extra limbs, extra arms, extra legs. When text and marks are unwanted, add: text, watermark, signature, username, logo. For framing and cropping problems, common words are: cropped, out of frame, cut off. Add one batch of words, generate a few images, and keep a word only after confirming it actually solves your problem — far steadier than piling on everything at once.

If you would rather not pick words from scratch every time, keep a general base list as a starting point that covers most common issues. A commonly suggested set is: lowres, blurry, bad anatomy, bad hands, extra fingers, extra limbs, deformed, poorly drawn face, poorly drawn hands, text, watermark, signature, logo, jpeg artifacts, low quality, worst quality.

Treat this list as a safety net, not an insurance policy: it lowers the probability of common errors and cannot replace solid positive prompt work on composition, lighting and subject. Use the whole set for everyday generation, then treat it as a living checklist: once a type of error disappears, its entry can go; when a new problem appears, add a new word for it. If details weaken or the style goes flat, start subtracting along the lines of the next section.

On top of the base list, just add a few targeted words per genre. Portrait photography usually takes: bad proportions, long neck, malformed limbs, cloned face. Anime and illustration generally take: extra digit, fewer digits, frame, comic style (when you want a single complete character image). Realistic photography usually takes: cartoon, anime, illustration, 3d render to suppress non-photo textures; skin-flaw entries are only for when you truly need clean-looking skin, not permanent residents of the list.

Keep scenario words restrained — three to five per genre is enough, and stacking everything defeats the point. Negative words can also backfire: putting cartoon into the negative makes genuine cartoon or illustration requests harder to produce. The same goes for common quality words, which can occasionally flatten the mood you want; when images come out dull and gray, try removing entries one by one to find the one dragging things down.

More negative words is not better; three side effects are common. First, conflicts: for example, negatively weighting opposite directions like low quality together with high detail leaves the model unsure which way to tighten the image. Second, dilution: the longer the list, the thinner the repelling force each word receives, and the few words that really matter may no longer hold the problem down. Third, degradation: with dozens of words, outputs often turn grayer, less saturated, less detailed and more conservative in composition, and occasionally opposite to what you intended.

A common suggestion is to keep negative prompts within ten to twenty entries, prioritizing words directly related to the current problem; the higher the CFG guidance scale, the stronger negative words act, which is all the more reason to stay lean. When you switch models or genres, re-check the list — a set that works well on one model does not necessarily transfer. After editing the list, generate a few images to compare and confirm you gained an improvement, not a side effect.

Finally, set the boundaries. A negative prompt lowers the chance of an element appearing but can hardly delete something already drawn: making a character take off their glasses, or turning a red outfit blue, is basically beyond negatives — hand those edits to inpainting, painting over the region and redrawing it separately, or simply rewrite the prompt. It also cannot raise the model's ceiling: if the base model draws hands poorly, piling on bad hands only eases the problem; it does not cure it.

The steadier approach is to treat negative prompts as the final fine-tuning step: write a clear positive prompt first (subject, composition, style, quality words), see where the default output falls short, then add three to five targeted negatives. Once positive-first, precise-negative becomes a habit, output stability usually improves noticeably, and the list keeps getting shorter over time.