Stable Diffusion negative prompt library
A negative prompt is the fastest way to clean up Stable Diffusion output. This library bundles curated, deduplicated presets for the most common image types — portraits, landscapes, anime and photorealism — plus a Universal quality set. Toggle the presets that match your scene, add any extra tokens, and copy a tidy negative prompt ready to paste into AUTOMATIC1111, Forge, ComfyUI, or any SD interface.
How negative prompts steer the model
During sampling, Stable Diffusion runs classifier-free guidance using both the positive and negative prompt. The model is pushed toward the positive concepts and simultaneously pushed away from the negative concepts. This is why extra fingers, deformed hands reliably reduces mangled anatomy, and blurry, low quality, jpeg artifacts sharpens detail — the model is directed away from those failure modes at each denoising step.
The effect is not removal. Negative prompts steer; they do not guarantee. A very strong artifact in the latent space may persist even with a corresponding negative token. But for the common surface-level quality issues — soft edges, color banding, flat lighting, distorted faces — targeted negative tokens are the most efficient fix.
Less is more — the common mistake
The instinct to paste every negative prompt you have ever seen into a single massive block is counterproductive. The model has a fixed capacity for each denoising step. An overstuffed negative prompt forces the model to steer away from dozens of concepts simultaneously, which dilutes the effect of each one and can introduce flatness, washed-out color, or reduced detail in areas that have nothing to do with your targets.
The reliable pattern is:
- Start with the Universal set (5–10 core quality tokens).
- Generate a test image.
- Identify a specific problem — extra finger, soft face, JPEG grain.
- Add only the token that addresses that specific problem.
- Repeat.
A 10-token negative prompt that precisely targets your model’s actual failure modes outperforms a 100-token dump of every negative ever documented.
Preset contents and when to use each
Universal: covers the most common cross-model quality degradations — blurry, low quality, watermark, text, cropped, out of frame, jpeg artifacts. Use this as the base for any generation.
Portrait: targets anatomy problems specific to human subjects — deformed face, extra fingers, bad hands, floating limbs, asymmetric eyes. Essential for any image with a person as the subject.
Landscape: suppresses the tiling, oversmoothed sky, and unnatural vignetting patterns common in environment shots. Include when the subject is scenery rather than a person.
Anime: tuned for anime/illustration checkpoints, suppressing the specific compression artifacts, off-model proportions, and style bleeding that appear on those models. Using photorealism negatives on an anime checkpoint tends to cancel the style you want.
Photorealism: targets plastic skin, HDR oversaturation, and AI’s tendency toward “stock photo” lighting on realistic checkpoints. Less useful on stylised models.
SDXL-specific note
SDXL generally needs shorter negative prompts than SD 1.5. It was trained on better-captioned data and produces fewer low-quality artifacts by default. If your SDXL output looks flat or desaturated after adding a full preset, trim the negative to a handful of core tokens and re-test. A universal set of 5–8 tokens is usually sufficient.