Negative prompts: what they fix and what is superstition

A negative prompt is the text the sampler steers away from. In classifier-free guidance it replaces the empty "unconditioned" prompt, so it really does change the result — just not always the way the long copied lists claim. Things that reliably help - Naming a concrete unwanted thing: text, watermark, signature, frame, border. - Naming the wrong medium when you want another: on a photo prompt, put painting, illustration, cartoon in the negative. - Short lists. Five targeted terms usually beat sixty. Things that mostly do not - Long anatomy lists (bad hands, extra fingers, mutated). On base SD 1.5 they have a weak, inconsistent effect. Hands get fixed by resolution, inpainting and ControlNet, not by asking nicely. - Contradictions. If the negative says "blurry" and the positive says "bokeh, shallow depth of field", you are fighting yourself. Negative embeddings Textual-inversion files such as EasyNegative were trained specifically to be used in the negative prompt. They work on the model family they were trained for and can shift style noticeably, so test with and without on the same seed. Distilled models LCM and other models run at CFG about 1-2 give the negative prompt very little influence, because guidance is so low. If a negative seems to do nothing on the Studio's fast engine, that is why. Method Same seed, same prompt: render with no negative, with your full copied list, and with three targeted words. Post all three. That comparison is more useful to the board than any list.