Negative Prompts: How to Tell AI What NOT to Generate

Half of professional prompt engineering is what you put in. The other half is what you keep out. Negative prompts — explicit instructions telling the model what to avoid — are the secret weapon behind clean, predictable AI images. This tutorial shows you when and how to use them, and what to put on your default exclusion list.

1. Introduction

A negative prompt is a second instruction sent alongside your main prompt, telling the diffusion model: "steer away from these things". Stable Diffusion has supported negative prompts since its earliest releases. Midjourney offers the --no parameter (e.g. --no text, watermark). DALL·E 3 does not have a formal negative-prompt field, but you can write exclusions in natural language ("with no people in the frame, no text on the bottle") and it generally listens.

Used well, negative prompts dramatically reduce the number of generations you need to throw away.

2. The Concept Explained

Imagine the diffusion process as a funnel. Your positive prompt pulls the model towards a region of latent space — say, "photorealistic portrait of a chef in a kitchen". But that region overlaps with many things you do not want: distorted hands, extra fingers, blurry text, garish colours, low-resolution noise. The negative prompt acts like a filter, narrowing the funnel before the final image is decoded.

Three types of useful negatives

Quality-defect negatives address known model failures: blurry, out of focus, low resolution, low quality, jpeg artefacts, pixelated, grainy, noise, deformed hands, extra fingers, six fingers, mangled face, asymmetric eyes, watermark, signature, logo, text, captions, ugly, disfigured.

Stylistic negatives push you out of unwanted style territory: cartoon, anime, illustration, painting, drawing, 3D render (when you want photorealism), or photorealistic, photographic, realistic skin (when you want illustration).

Content negatives remove specific elements: people, crowds, cars, modern technology, text, signs, jewellery, hats. Use these to keep scenes uncluttered or period-accurate.

3. The Problem Without Negative Prompts

portrait of a young woman holding a coffee cup,
in a warm Italian cafe

4. The Solution

POSITIVE: portrait of a young woman in her late twenties,
gentle half-smile, holding a small espresso cup with both
hands, sitting in a warm Italian cafe with soft morning
window light.

Style: photorealistic, 85mm portrait lens, shallow depth
of field, Kodak Portra 400 aesthetic.

NEGATIVE (Stable Diffusion field, or Midjourney --no):
deformed hands, extra fingers, mangled fingers, text on cup,
watermark, logo, brand text, passers-by, blurry, low quality,
oversaturated, cartoon, illustration, painting

--ar 4:5 --v 6 --style raw --no text, deformed hands,
extra fingers, watermark

5. Step-by-Step Breakdown

1. Build a default exclusion list. Keep a saved list of quality-defect negatives — deformed hands, extra fingers, watermark, text, blurry, low quality — that you paste into almost every prompt.
2. Add stylistic negatives. If you want photorealism, exclude cartoon, anime, illustration, painting. If you want illustration, exclude photorealistic, photographic.
3. Add prompt-specific content negatives. If you want a quiet street, exclude people, crowds, cars. If you want a clean product shot, exclude props, hands, fingers.
4. Adapt to the platform. Stable Diffusion has a dedicated negative field. Midjourney uses --no element1, element2. DALL·E 3 accepts plain-language exclusions inside the prompt.
5. Iterate by adding, not rewriting. When a generation has a specific defect (a hat you didn't ask for, a phone in someone's hand), add that single item to your negative list and re-run.

Tip: Do not stuff your negative prompt with hundreds of words. Twelve to twenty well-chosen negatives outperform a sprawling 200-word block. Overly long negatives can paradoxically reintroduce the very concepts they mention by giving them more weight.

6. Practice Exercises

7. Key Takeaways