# Negative prompts explained: when to use them, and how Canonical: https://romeoapps.app/en/larpgpt/negative-prompts-explained/ Author: Roméo Gambino · Published: 2026-09-10 · Updated: 2026-09-10 · App: LarpGPT (https://romeoapps.app/en/larpgpt/) What a negative prompt is, which image tools document one, how Midjourney's --no and Stable Diffusion's negative_prompt work, and why to use them sparingly. ## What a negative prompt is A negative prompt is a list of things you do not want in a generated image, given to the tool separately from the description of what you do want. Instead of writing a street at night with no people, you describe the street and put people in the negative list. The idea is simple, but tools handle it very differently, and some do not offer a separate negative field at all. Knowing which situation you are in saves a lot of failed attempts. ## Why writing no inside the prompt often fails Image models respond to the words present in a prompt. Mentioning something, even to exclude it, can pull it into the picture. Midjourney's documentation gives exactly this kind of example as a bad prompt: a still life painting described as without any fruit, or with a polite request not to add fruit. The recommended alternative is to move the unwanted elements into the dedicated parameter. The same logic applies elsewhere: a sentence full of exclusions keeps the excluded things in the model's attention. ## Midjourney: the --no parameter Midjourney documents a --no parameter, added at the end of the prompt and followed by the thing or list of things you do not want, separated by commas. Its own example is "still life gouache painting --no fruit, apple, pear". The documentation also explains that using --no is the same as giving part of a multi-prompt a weight of minus 0.5, which is a useful reminder that it reduces something rather than guaranteeing its absence. ## The trap in Midjourney's --no The same documentation warns that Midjourney's moderation reads every word after --no independently. A phrase such as modern clothing is read as two separate words, and the word clothing on its own can trigger a warning. In cases like that, the documentation advises including the kind of clothing you do want in the main prompt instead of using --no. More generally, keep each excluded item to a single, unambiguous word where you can. ## Stable Diffusion: the negative_prompt argument In the Hugging Face diffusers library, the Stable Diffusion text-to-image pipeline has an optional negative_prompt argument, described in its documentation as the prompt or prompts that guide what not to include in image generation. The documentation adds a detail worth knowing: the negative prompt is ignored when guidance is not used, meaning when the guidance scale is below one. ## ChatGPT and OpenAI's image models OpenAI's image generation documentation, at the time of writing, does not describe a separate negative prompt field for its GPT Image models. With conversational tools you can still write exclusions in a sentence, but the most reliable approach is the one that works everywhere: describe what should be there. If you do not want people in a street, describe an empty street at dawn with the shops still closed. That gives the model a positive picture to build rather than a list of things to avoid. ## Describe the alternative first Before reaching for a negative, ask what you want in place of the unwanted element. Clutter on a desk becomes a clear desk with one notebook and a lamp. A busy background becomes a plain concrete wall. Harsh shadows become soft overcast light. Almost every negative has a positive version, and the positive version usually works better, because it constrains the image instead of leaving a gap the model fills with something else. ## When a negative prompt is the right tool Negatives are most useful for things that tend to appear uninvited and have no natural positive opposite: text, lettering, watermarks, logos, borders or frames. They also help when a generator keeps adding a recurring element, such as people in a landscape or lens flare in a sunset, after you have already described the scene clearly. In those cases a short exclusion is simpler than rewriting the whole scene. ## Keep the list short Long negative lists are a common habit copied from shared prompts, often filled with generic words about quality. They rarely help and can make results harder to predict, because every word in the list pushes the image in some direction. Start with no negatives at all. Add one only when a specific unwanted element keeps appearing, check whether it worked, and remove it again if it did not. Two or three precise items are easier to reason about than twenty vague ones. ## Use nouns, not qualities Negatives work best with concrete things: text, people, cars, trees, watermark. Abstract qualities such as ugly, bad or low quality are hard for a model to interpret, and they say nothing about what you actually want to change. If an image looks wrong, work out what is physically wrong with it, whether the hands, the reflections or an extra object, and address that directly, either in the description or with a precise exclusion. ## Test one negative at a time Treat a negative like any other change to a prompt. Generate the image without it, then add a single exclusion and generate again, keeping everything else the same. If the unwanted element disappears and nothing else breaks, keep it. If the image changes in ways you did not expect, the negative was doing more than you intended. This habit also teaches you which exclusions are worth saving for later and which were superstition. ## Negatives do not replace checking No exclusion guarantees an absence. Midjourney describes --no as a reduction in weight, and diffusers describes negative_prompt as guidance. Look at every image before using it: text can still appear on signs, a logo-like shape can appear on clothing, and a figure can appear in the distance. If an image might be taken for a real photograph, label it as generated. ## Where LarpGPT fits LarpGPT is a library of photo prompts for travel, architecture, cars and lifestyle scenes, written mostly as positive descriptions of subject, composition, lighting and lens, which reduces the need for long negative lists. You copy a prompt into the image generator you already use and add any exclusions that tool supports. LarpGPT does not generate images, and results vary between attempts. ## Questions ### What is a negative prompt? A list of things you do not want in a generated image, given separately from the main description. Midjourney uses the --no parameter, and Stable Diffusion in the diffusers library uses a negative_prompt argument. ### Does ChatGPT support negative prompts? OpenAI's image generation documentation does not describe a separate negative prompt field at the time of writing. Describing what you want instead, such as an empty street rather than a street without people, is the more reliable approach. ### Why does writing without something add it to the image? Models respond to the words present in the prompt, so naming an object, even to exclude it, can bring it into the image. Midjourney's documentation lists this as a bad prompt pattern and recommends the --no parameter instead. ### How many items should a negative prompt contain? As few as possible. Start with none, add one precise item when an unwanted element keeps appearing, and remove it if it does not help. Long generic lists make results harder to predict. ## Further reading - Midjourney documentation: No (archived copy, July 2026) — https://web.archive.org/web/20260731111101/https://docs.midjourney.com/hc/en-us/articles/32173351982093-No - Hugging Face diffusers: Stable Diffusion text-to-image pipeline — https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/text2img - Hugging Face diffusers: Prompting and weighted prompts — https://huggingface.co/docs/diffusers/using-diffusers/weighted_prompts - OpenAI API documentation: Image generation — https://platform.openai.com/docs/guides/image-generation