Lesson 3. Negative prompt: how to tell AI what shouldn't be in the image
AI images often contain deformed hands, unnatural faces, watermarks, or low quality. The negative prompt is the main tool against these problems. One good negative prompt block raises image quality by several levels.
Topic breakdown
A negative prompt is an instruction to the AI about what should not be in the image. The main prompt says 'what is needed', the negative prompt says 'what is not needed'.
Stable Diffusion and most Leonardo.ai interfaces have a separate field for negative prompts. Midjourney uses the parameter '--no [element]'. In DALL·E, negative concepts are added to the main prompt via 'without X'.
There is a standard basic negative prompt block: low quality, anatomical errors, watermarks, extra joints, etc. It can be copied and used in every project.
Additionally, you need to block specific elements based on context: in a portrait — 'glasses' (if glasses are not needed), in architecture — 'people, cars' (if only the building is needed).
What you'll learn
- Understand the principle of negative prompts
- Know negative prompt syntax for different models
- Master standard quality blocks and use them constantly
- Select specific negative elements depending on the context
Deep dive
A negative prompt should be seen not as magic that creates a beautiful picture, but as a cleaning layer. It stabilizes the scene created by the main prompt and reduces the number of common artifacts.
It's best to store a universal negative prompt block separately from contextual blocks. For example, a general quality block for all tasks, an anatomy block for portraits, and a text/logo block for product images.
If a negative prompt becomes too long, it starts competing with the main prompt. Therefore, keep only the 6-10 most frequent problems in the constant part, and add the rest according to the situation.
Lesson plan
How negative prompt works
Diffusion models consider the main and negative prompts simultaneously during generation. Unwanted elements are constantly 'pushed away'.
Syntax by model
SD/Leonardo: separate negative prompt field. Midjourney: '--no element1, element2' at the end of the prompt. DALL·E: 'without X, avoid Y' in the main prompt.
Standard basic block
Most professional users use the same basic negative prompt block. Save it and use it as a starting point in every new project.
Specific blocking strategy
For each project, add specific elements to the base block: for product photos — 'people', for architecture — 'modern elements', for portraits — 'artifacts, noise'.
Weak vs strong prompt
beautiful portrait of a man, professional photo (negative: nothing)
portrait of a man in his 40s, professional headshot, soft studio lighting, sharp focus, photorealistic --no blurry, bad anatomy, extra fingers, deformed face, watermark, low quality, pixelated, noise
The negative prompt blocks anatomical errors, low quality, and artifacts. This is especially noticeable in images of hands and faces.
Ready visual prompt template
Copy and adaptBasic negative prompt block (SD/Leonardo): blurry, low quality, low resolution, bad anatomy, deformed, ugly, extra limbs, missing fingers, watermark, text, logo, cropped, out of frame, duplicate
Why it works
'blurry, low quality, low resolution' — essential quality blocks, always add them.
'bad anatomy, deformed, extra limbs, missing fingers' — mandatory for portraits and images of people.
'watermark, text, logo, signature' — if a clean image is needed.
'cropped, out of frame' — when the object must be fully visible.
Specific blocking: 'glasses' (no glasses), 'cars' (no transport), 'modern buildings' (historical scene).
Practice
- Write a portrait prompt: 'young woman, outdoor portrait, natural light, photography'.
- First generate without a negative prompt and evaluate the results.
- Now add a negative prompt: 'blurry, bad anatomy, deformed, extra fingers, watermark, low quality'.
- Compare both results: pay attention to hands, face, and background quality.
Checklist
Common mistakes
- using 'good' words in a negative prompt: do not write 'no blur', just write 'blur'
- too short negative prompt: just 'low quality' is not enough
- adding elements from the main prompt into the negative prompt (mutual blocking)
- searching for a separate field in Midjourney instead of using the '--no' parameter
Lesson FAQ
How long should a negative prompt be?
20-40 keywords is optimal. An overly long negative prompt can reduce the strength of the main prompt. Focus on the most important elements.
How do I use a negative prompt in DALL·E?
DALL·E doesn't have a special field for a negative prompt. Instead, add 'without X', 'no Y', 'avoid Z' to the main prompt. For example: 'a clean product shot, white background, no shadows, no people'.