Imagery
The blocklist: the fastest quality gain in any prompt
What you don't ask for matters as much as what you do. A well-built blocklist improves an image more than twenty further adjectives.
By: Vadász MárkPrompting video models (Seedance, Kling, Veo)
Most people keep adding to the prompt until the picture comes right. That is the wrong direction: the more adjectives you pile on, the more the composition falls apart, because the model has to satisfy all of them at once.
The quicker route is to say what to leave out. That doesn't constrain the model, it frees it: the attention it saves goes into what you actually asked for.
Four groups you always need
- Technical faultsBlur, noise, compression artefacts, extra limbs, fused fingers. These are not matters of taste: nobody wants them.
- Style driftAdvertising gloss, perfect skin, excessive symmetry. This is the model's resting state, so it has to be ruled out explicitly.
- Text and watermarksGenerators like putting text on images, because their training data was full of it. One stray logo makes an otherwise good frame unusable.
- Whatever doesn't belong in your sceneThis is the most useful group and the one nobody can write for you: modern objects in a period scene, a smiling face in a grim moment, sunlight in a night shot.
Base list: write it once, reuse it on every project
- 1. Technical faults
blurry, extra fingers - 2. Style drift
plastic skin, airbrushed - 3. Text and watermarks
watermark, text overlay
Written per scene
- 4. Whatever doesn't belong in your scene
modern cars, neon signs
A list to start from
blurry, low resolution, jpeg artifacts, extra fingers, deformed hands, watermark, text overlay, signature, oversaturated, plastic skin, airbrushed, perfect symmetry, stock photo look
Paste this in as your base and extend it with whatever is specific to the scene. Take out anything that doesn't apply: a long list of prohibitions distracts the model exactly as much as a long list of requests.
Video needs a different list
In moving image the failures show up differently. The hands are not the problem; the problem is a face that morphs mid-shot, a camera move that judders, or a background that boils.
| You block in stills | You block in video |
|---|---|
| extra fingers, deformed hands | morphing face, shifting clothes |
| watermark, on-screen text | juddering camera moves, abrupt jump cuts |
| over-sharpening | flickering background, boiling texture |
What not to block
Don't block „bad quality" or „ugly" in general. Those mean nothing to a model and usually change nothing at all. A prohibition works when it names something SPECIFIC, exactly as a request does.
A blocklist isn't protection against mistakes. It is the direction the model is not allowed to take.
Sources
- Midjourney documentation: No (the --no parameter)
- Hugging Face Diffusers: Prompting, guidance and negative_prompt
- AUTOMATIC1111 Stable Diffusion WebUI wiki: Negative prompt
- Ho, J. and Salimans, T. (2022): Classifier-Free Diffusion Guidance, arXiv
Common questions
What is a negative list in an AI prompt and why use one?
A negative list tells the model what to leave out of the image. That does not restrict it, it frees it: the attention it saves goes into what you actually asked for. A well-built list improves the image more than twenty extra adjectives, which tend to make the composition fall apart.
What should I put in a negative prompt?
Think of four groups: technical flaws (blur, noise, extra fingers), a slide back to the default style (perfect skin, advertising light), text and watermarks, and anything that does not belong in your scene. You build the first three once and reuse them. The fourth you write for each scene.
Is a negative list for video different from one for images?
Yes, because the flaws show up differently. In an image the trouble is a deformed hand or a watermark. In video it is a face that changes during the scene, clothes that shift, a jerky camera and a flickering background. So a video list names these motion-related problems.
Does adding ugly or bad quality to a negative prompt help?
Usually not. To the model, general words like these mean nothing, so they rarely change anything. A prohibition works when it names something specific, such as blur, a watermark or extra fingers, in the same way that a request only works when it is specific.