Negative prompt

Most diffusion models use a technique called classifier-free guidance. At each step the model makes two predictions: one conditioned on your prompt and one on a reference with no instruction. The final step moves away from the reference and toward your prompt, and the guidance setting says how far. A negative prompt replaces that empty reference with a description of what you do not want, so the image is pushed away from it as well as toward what you asked for. Software documentation sometimes calls the second condition the unconditional or null prompt, which explains why replacing it is possible at all.

That explains both its power and its limits. Models that run without guidance, which is common with very fast ones trained to work in a few steps, have no reference to replace, so they often ignore a negative prompt altogether. Even where it works, it is a nudge and not a guarantee: naming something you want to avoid can sometimes make it more present, much as being told not to think of a pink elephant does, and a long list of vague words rarely does more than a short one. Weighted syntax such as parentheses around a word is a convention of certain interfaces, not a universal rule.

Good habits are simple. Describe the main subject fully in the main prompt, and keep the negative prompt for a few recurring flaws: blur, jpeg artifacts, distorted hands, duplicated faces, cut-off framing, stray captions or a watermark you do not want. For a product shot on a clean backdrop, reflections, dust, harsh shadows and a tilted horizon are better entries than words such as ugly or bad, which describe no visual feature the network can steer away from. Change one thing at a time and compare results with the same seed, so you can attribute each difference to its cause, like a photographer turning a single dial between exposures, and keep a short log of winning combinations. Beware of contradictions as well: asking for a vivid sunset while excluding warm tones leaves the network torn between two orders. When a flaw persists, rewording the positive prompt, for example asking for a sharp, close-up photograph, often works better than adding to the negative list.

The Studio has a negative prompt field of up to 500 characters, and the API accepts an optional negative_prompt of the same length, used by the models that support it. Support depends on the model, so if a field seems to change nothing, try another model before changing your wording. The image generator page describes the rest of the Studio: eight formats and fifteen styles, with the default model Z-Image Turbo, an open model included in every plan and quick enough to make comparing two wordings cheap.

One detail worth knowing, stated on the transparency page: the Studio settings, the negative prompt included, are kept in plain text in your browser's local storage, with no conversation content. Your generated images themselves are stored encrypted in the browser, and the hard limits of NeverNo apply to every request whatever a negative prompt says.