Earngenix Logo
Skip to main content

Glossary · ComfyUI Basics

What Is a Negative Prompt in ComfyUI?

By Earngenix Team ·

⚡ Quick Answer

A negative prompt is text describing what you don't want in the image — like "blurry" or "extra fingers" — fed into KSampler alongside your normal (positive) prompt so the sampler steers away from those results.

KSampler takes two separate conditioning inputs: one for what you want (positive) and one for what to avoid (negative). During each denoising step, the model weighs the image toward the positive conditioning and away from the negative conditioning — the negative prompt doesn't just get ignored, it actively pushes the result in the opposite direction, and the CFG scale controls how strongly.

Where You'll See It

A second CLIP Text Encode (Prompt) node, wired into KSampler's negative input — visually identical to the positive prompt node, just connected to a different socket and containing different text.

Quick Example

A common general-purpose negative prompt reads something like: "blurry, low quality, extra fingers, watermark, text, deformed hands" — targeting the specific flaws you're actually seeing tends to work better than a long generic list.

Newer models like Flux.1 Dev were trained without classifier-free guidance and often ignore the negative prompt entirely at the default cfg value — check your specific checkpoint's documentation if negative prompts don't seem to be doing anything.

Common Mix-Up

Negative prompt vs. low prompt weight: lowering a word's weight in your positive prompt (like "(cat:0.5)") just reduces how much that word matters. A negative prompt actively steers away from a concept — the two aren't interchangeable ways of removing something.

Frequently Asked Questions

No — you can leave it blank or use an empty CLIP Text Encode node and generation still works fine. It’s an optional tool for steering away from specific unwanted results, not a required part of every workflow.

Not necessarily. Piling on dozens of generic terms like “bad quality, worst quality, low resolution” has diminishing returns and can occasionally fight your positive prompt. Targeted terms for the specific problem you’re seeing usually work better than a huge generic list.

Yes — negative embeddings trained specifically to represent common flaws (blurriness, bad anatomy, and similar) are a popular shortcut, and you call them the same way as any other embedding, by name, inside the negative CLIP Text Encode node.

See It In Action

Ready to write your first prompt pair?

Our text-to-image tutorial covers positive and negative prompting from scratch.

Published: 2026-09-04 · Last updated: 2026-09-04

Discussion

Join the discussion

Sign in to leave a comment or reply

💬

No comments yet

Be the first to share your thoughts!