What is a negative prompt?

A negative prompt is not a filter applied afterwards, and it is not the model ignoring words. It is a second full prompt encoded exactly like the first one. At every denoising step the model predicts two directions, one toward your prompt and one toward your negative prompt, and then moves toward the first while actively moving away from the second.

That mechanism explains what negative prompts are good at and what they are not. They work well for pervasive qualities that affect the whole image: blur, low quality, oversaturation, watermarks, extra limbs, text. They work poorly for removing a specific object from a specific place, because the push is applied globally rather than to a region. If you want a landscape with no house in it, describing the landscape well beats writing "house" in the negative.

The common mistake is length. Long copy-pasted negative prompts full of dozens of terms are widely shared, and they mostly dilute each other: the further the negative direction points, the less precisely it points anywhere. Three to six specific terms aimed at problems you are actually seeing will do more than forty generic ones.

On Dreamify you enable advanced features in the generator to get the negative prompt field, alongside the scheduler, inference steps and seed. A reasonable default for photographic work is something like "blurry, low quality, watermark, text, distorted hands", trimmed to whatever you keep having trouble with.