피파 한 줄 정리: Negative prompt는 모델별 지원 차이가 커. SD에선 필수, FLUX에선 제한적, Midjourney·DALL-E에선 다른 메커니즘. '키친 싱크 negative'는 옛날 SD 1.5 유산.
Negative prompting is one of the most misunderstood features in image generation. It sounds simple — "tell the model what you don't want" — but the reality is nuanced and model-dependent.
How Negative Prompts Work (Technically)
Remember classifier-free guidance? The model compares conditional (with text) vs. unconditional (without text) predictions. Negative prompting replaces the unconditional prediction with a negative-conditional prediction:
Standard CFG: Output = Unconditional + Scale × (Conditional − Unconditional) With negative prompt: Output = Negative_Conditional + Scale × (Conditional − Negative_Conditional) The model pushes AWAY from the negative prompt and TOWARD the positive prompt.
So a negative prompt doesn't just "block" unwanted elements — it defines what the model actively steers away from. The generation is pushed in the opposite direction from whatever the negative prompt describes.
Model Support Varies Widely
| Model | Negative Prompt Support | Notes |
|---|---|---|
| SD 1.5 / SDXL | ✅ Full support | Very effective; widely used |
| SD 3.5 | ✅ Supported | Works but less critical than SD 1.5 |
| FLUX | ⚠️ Limited | Supported in some implementations; less effective than in SD |
| Midjourney | ❌ No traditional negatives | Uses --no parameter instead (simpler) |
| DALL-E 3 | ❌ Not supported | Relies on prompt rewriting instead |
When Negative Prompts Help
- Removing consistent unwanted artifacts: "blurry, low quality, watermark" in the negative prompt can clean up outputs for models that support it.
- Excluding specific elements: "no text, no border, no frame" can help avoid unwanted overlays.
- Countering model biases: If a model tends to produce overly saturated images, "oversaturated, HDR" in the negative can tone things down.
When Negative Prompts Fail or Backfire
Positive: "a dog in a park" / Negative: "cat"
Mentioning "cat" in any context — even negatively — activates cat-related patterns. The model sometimes produces cat-like features BECAUSE you mentioned cats. For simple exclusions, it's better to just not mention the unwanted element at all.
Negative: "ugly, deformed, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, mutated hands, fused fingers, too many fingers, long neck, username, watermark, signature"
Negative: "blurry, low quality, watermark"
The kitchen-sink negative was essential for SD 1.5 because that model frequently produced anatomical errors. For modern models that rarely produce these issues, massive negative prompts can actually constrain the output space too much, reducing diversity and creativity.
- Negative prompts steer generation away from specified patterns — they don't "block" specific pixels.
- Support varies by model: essential for SD 1.5, limited for FLUX, absent for Midjourney/DALL-E 3.
- Mentioning something in the negative can paradoxically activate it — sometimes it's better to just not mention it.
- Keep negatives short and focused. Kitchen-sink negatives from the SD 1.5 era are usually unnecessary for modern models.