This entry is about negative prompts—the part that tells the model what not to include: how they’re categorized, where to place them, and how many actually work. It’s for creators whose positive prompts are already solid, but whose output keeps failing in the same spot.
Four Categories of Negative Words
Categorized by failure type—pick what you need, don’t copy everything:
- Quality: blurry, low quality, artifacts → fixes soft, noisy, dirty frames.
- Anatomy: extra fingers, deformed hands, distorted face → fixes broken limbs and bodies; nearly essential for close-ups and hand-action shots.
- Motion: frozen, static image, jittery motion → fixes “a breathing photo” or erratic speed changes; unique to video models.
- Consistency/style drift: face changing, style drift, watermark, subtitles → fixes subjects that change appearance mid-shot, or stray text and logos appearing in frame.
Placement
- Models with a dedicated negative field (a separate negative prompt input box): put them there—don’t mix them into the positive prompt.
- No dedicated field: leave a blank line at the end of the prompt, then start an
AVOID:section or use a--noparameter, e.g.:AVOID: blurry, extra fingers, static image, watermark. Which syntax to use depends on that model’s official docs. - How strongly each model parses a trailing AVOID section varies: unverified (to be filled in after real-world testing).
Keep It Lean
- Negative words are a fixing tool, not a lucky charm: you can write the first draft with zero negative words; run it once, then add words based on what actually broke.
- Keep it to 3–8 words; beyond 10, they dilute each other and each word’s pull weakens.
- The most effective negative words come from the last failure: if the previous run gave six fingers, add extra fingers; if the frame didn’t move, add static image. This is exactly how this works with the [prompt iteration method] entry—change only one thing per round, and adding/removing from the negative section counts as one change, making attribution easy.
- Clear it out when switching projects: the negative section from the last project doesn’t carry over automatically. Start from zero and rebuild based on the new project’s failures.
Common Mistakes
- Copy-pasting 30 “all-purpose” negative words verbatim → wastes space, dilutes each other, and the two or three you actually need lose their grip
- Accidentally putting what you want into the negative section (wanting slow motion but casually writing slow motion in AVOID) → negating your own positive request
- Negative and positive directly contradicting each other (positive says static camera, negative says static) → conflicting instructions, output swings randomly
- Not checking the docs and using the same negative format across all models → if the syntax isn’t recognized, the negative section may be read as body text, turning negative words into summoning words instead (actual behavior varies by model: unverified)