The most common mental model for negative prompts is wrong. Most people treat them like a blacklist: list the things you don't want, and the model removes them. This model predicts that more negative terms produce better results, that listing something in the negative prompt guarantees its absence, and that negative prompts work reliably regardless of what the positive prompt says. None of these predictions are accurate.
The reason negative prompts sometimes seem to work like a blacklist is that the underlying mechanism produces similar outcomes in simple cases. But once you try to use negative prompts to solve a specific problem — removing figures that keep appearing, suppressing a particular color cast, preventing a recurring artifact — the blacklist model fails to explain why some approaches work and others don't. Understanding the actual mechanism produces much better results.
What Negative Prompts Actually Do
Stable Diffusion generates images through a process called classifier-free guidance. During generation, the model runs two simultaneous conditional passes: one conditioned on your positive prompt (generating toward what you described) and one unconditioned (or conditioned on the negative prompt). The final output is a weighted combination of these two directions — moving toward the positive prompt's region of the latent space and away from the negative prompt's region.
The key implication: negative prompts work by moving generation away from a direction in a high-dimensional space. They are not filters applied to the output. They are directional signals applied during the generation process itself. This distinction has several practical consequences:
Negative prompts affect the entire image, not just the targeted region. There is no spatial targeting in a standard negative prompt. "Person" in the negative prompt pushes the entire generation away from the person-associated region of the latent space, which affects not just whether a person appears but potentially the rendering style, atmosphere, and detail distribution of the entire image.
Negative prompt strength is relative, not absolute. If the positive prompt has very strong figure-association pressure — from explicit figure terms or from atmospheric vocabulary that strongly implies human presence — the negative prompt's push away from figure content can be insufficient to overcome that pressure. The positive and negative signals are in competition. A strong positive pull toward figures can overpower a moderate negative signal against them.
Listing more terms does not linearly increase suppression strength. The negative prompt is processed through the same text encoder as the positive prompt. Adding many terms dilutes the attention weight given to any individual term. A negative prompt with 80 words may suppress any single concept less effectively than a focused negative prompt with 10 words targeting the same concept.
What CFG Scale Does to Negative Prompts
The CFG (Classifier-Free Guidance) scale parameter directly controls how strongly both the positive and negative prompts influence the generation. Higher CFG means the model follows both prompts more aggressively. This means:
- At low CFG (below 5): negative prompts have minimal effect. The guidance signal is weak, and the model's output distribution is relatively free.
- At medium CFG (6 to 9): negative prompts have meaningful influence. This is the functional range for most negative prompt applications.
- At high CFG (above 12): negative prompts have strong influence, but the overall generation becomes saturated, over-contrasted, and prone to artifacts as both positive and negative directions are amplified beyond the model's comfort zone.
If your negative prompts seem to have no effect, the first thing to check is whether CFG is set too low. If your outputs look harsh and over-processed, the first thing to check is whether CFG is too high relative to the combined weight of your positive and negative prompt content.
How to Structure Negative Prompts for Specific Problems
For quality and artifact suppression
This is the most common negative prompt use case: suppressing common AI generation artifacts and quality degradation patterns. The standard quality-focused negative prompt terms work because these concepts are associated in training data with lower-quality image characteristics.
A functional quality-suppression negative prompt for most SD 1.5-based models:
bad quality, worst quality, low quality, normal quality, jpeg artifacts, blurry, out of focus, watermark, signature, text, username, extra limbs, deformed, ugly, low resolution, grainy
This should be treated as a base layer — it addresses the most common quality failure modes and can be appended to any generation without causing problems. Keep it focused on well-established quality terms rather than expanding it with every artifact you've ever seen.
For suppressing figures in empty scene generation
As discussed in more detail in the empty atmospheric scenes article: figure suppression requires a layered negative prompt approach. The base quality negative prompt is insufficient for this purpose. You need terms targeting human presence specifically:
person, people, human, figure, man, woman, character, silhouette, portrait, face, body, hands, arms, standing figure, seated figure, walking figure
This should be used in combination with positive prompt environmental framing (describing the scene from an architectural rather than emotional perspective). The negative prompt suppresses figure association; the positive prompt removes the figure-implying pressure from the generation direction.
For suppressing a specific color cast
Certain models, fine-tunes, and VAE combinations produce consistent color casts — a warm orange tint, a green shift in shadows, an overall desaturated gray haze. The negative prompt can push against these tendencies:
For warm orange cast: orange tint, warm color cast, oversaturated oranges, sepia toning
For green shifts: green tint, green color cast, chlorine color, greenish shadows
For gray desaturation: desaturated, gray, washed out, flat colors, muted
Note: negative prompts for color correction have limited precision and are less reliable than correct VAE selection and CFG calibration for addressing systematic color issues. They are a nudge, not a correction.
For suppressing specific stylistic tendencies
Models sometimes produce stylistic tendencies that conflict with the intended aesthetic. Suppressing these requires identifying which concepts in the model's training associate with the unwanted tendency:
For overly plastic skin or toy-like surfaces: plastic, shiny, glossy, oversmoothed, doll-like
For excessive lens flares and cheap photographic effects: lens flare, bloom, chromatic aberration, vignette
For overcrowded compositions: busy, cluttered, too many objects, chaotic
For overused AI aesthetic clichés: trending on artstation, hyperdetailed, 8k uhd (counterintuitively, these quality-boosting terms sometimes produce over-rendered, over-processed aesthetics in certain models)
