Negative prompts: telling the AI what to avoid

A negative prompt is a blocklist for the AI — it describes what you want to avoid. One of the simplest ways to improve generation quality.

When to use this in your workflow

When to use:

On almost every generation. A basic negative prompt reduces common artifacts (watermarks, text, anatomical errors) across the board.
When you see recurring artifacts. Same problem keeps appearing? Add those specific terms: “soft focus, text overlay, oversaturated.”
To exclude a specific style. “Cartoon, anime, 3D render” keeps output photorealistic. “Photorealistic, photograph” pushes toward illustration.

When NOT to use:

Model doesn’t support it. Newer architectures (including most Flux models) ignore negative prompts entirely. If modelBridge doesn’t show the field, the model doesn’t support it.
Don’t overload it. A negative prompt with 50 terms fights itself and produces bland, generic output. Keep it under 10–15 terms.

Starter negative prompt

Copy this as your baseline:

blurry, watermark, text, low quality, deformed, disfigured, extra limbs, bad anatomy

Add model-specific or project-specific terms as needed.

How it works in modelBridge

The negative prompt field appears in advanced settings of models that support it. modelBridge’s dynamic UI only renders it when the model’s schema includes it — if the field isn’t there, the model doesn’t support negatives.

No cost impact. Negative prompts are processed as part of the same request. Price is identical whether you use one or not.
Works as a pair with your main prompt. Your prompt pulls toward what you describe. The negative prompt pushes away from what you list. Both well-written = clearer target = fewer wasted generations.
Interacts with guidance scale. High CFG amplifies both prompts. Strong negative + high CFG can overcorrect and produce flat output. Keep CFG at default when using aggressive negatives. See The 4 parameters.

Recommended settings

Baseline: “blurry, watermark, text, low quality, deformed, bad anatomy” — covers the most common artifacts.
For photorealism: Add “cartoon, anime, 3D render, illustration, painting.”
For illustration: Add “photorealistic, photograph, camera.”
Max terms: 10–15. Beyond that, diminishing returns.
Keep CFG at default when using a strong negative prompt.

Common mistakes

Writing “no ugly background” instead of “ugly background.” The model reads words, not intent. “No trees” often causes trees. Put “trees” in the negative field instead.
Pasting a 50-word blocklist into every generation. The model tries to avoid everything at once and produces bland output.
Adding terms “just in case.” Only add terms that target problems you’ve actually seen.
Using negatives on Flux models. Flux doesn’t support negative prompts. If the field isn’t showing, don’t try to force it.

Quick answers

What’s a good starting negative prompt? “Blurry, watermark, text, low quality, deformed, bad anatomy.” Covers the most common artifacts without being restrictive.

Do all models support negative prompts? No. Many newer models (including Flux) don’t use them. If modelBridge doesn’t show the field, the model doesn’t support it.

Can a negative prompt make output worse? Yes, if it’s too long or contradicts your main prompt. Keep it focused on specific problems.

Should I use the same negative prompt every time? A standard set works as a baseline. Add project-specific terms as needed.

What’s the difference between “no trees” in the prompt vs. “trees” in the negative prompt? “No trees” in the main prompt often causes the model to generate trees (it reads “trees” as a keyword). “Trees” in the negative prompt is more reliable for exclusion.