Negative Prompts, What They Remove and What They Don't

The results for negative prompts are a strange set. A Reddit thread asking what people have found works, a forum post from a CAD product's community, a couple of listicles offering 150 and 200 terms respectively, one vendor's documentation, and then, sitting at the bottom of page one, an academic paper that actually explains the mechanism. Nobody above it has read it, as far as I can tell.

The answer most people need is short. Negative prompts remove specific things from a picture. They do not add quality, they are not a slider, and on at least one widely used implementation they are silently ignored entirely depending on another setting. Piling on forty terms to make an image "better" is the single most common waste of effort in image prompting, and the vendor whose documentation ranks on this very page tells you not to do it.

Below is what each system actually supports, what the mechanism turns out to be, and how to test whether yours honours them at all.

Not Every System Has The Field

First problem. People talk about negative prompts as though they're a universal feature, and they're not, and the word "prompt" in the name hides that the good implementations are a separate input rather than something you type into the same box.

Here's what I could confirm from documentation loaded today. Where a page didn't mention it, I've said so rather than assuming absence.

Two implications fall straight out of that. If you're using one of the systems in the bottom three rows, every "negative prompt" tutorial written for Stable Diffusion is describing a control you don't have, and the equivalent move is writing exclusions into your ordinary sentence, which behaves differently and less predictably.

The other implication is subtler. On systems without the field, an exclusion competes with your description for the model's attention rather than sitting outside it, so a long list of things you don't want is spending the same finite budget as the things you do want. That trade is invisible and it's real.

The Setting That Switches Them Off

This is the fact I'd most like people to know, and it's sitting in plain sight in the diffusers pipeline reference.

The negative_prompt argument is documented as "Ignored when not using guidance", where guidance is enabled when guidance_scale is above 1. The default guidance_scale in that pipeline is 7.5, so most people are fine, but plenty of the fast sampling setups and distilled models people run now operate at a guidance scale of 1 or below, and in those configurations your carefully assembled negative prompt does nothing at all.

Nothing. Not "less". It's not passed through.

I'd guess that accounts for a decent share of the "negative prompts don't work for me" posts, and it's the sort of thing that never appears in a list of 200 recommended terms, because the list genre assumes the mechanism works and only argues about vocabulary.

Worth checking your own setup before you spend an evening tuning wording.

What The Paper Found

The academic result on page one is Ban, Wang, Zhou, Cheng, Gong and Hsieh, "Understanding the Impact of Negative Prompts: When and How Do They Take Effect?", submitted in June 2024. It's an empirical study of the mechanism in models like Stable Diffusion and it names two behaviours.

The first is a delayed effect. Their finding is that the impact of a negative prompt is observed after the positive prompt has rendered the corresponding content. The thing has to start existing before it gets removed.

The second is deletion through neutralisation, where the negative prompt cancels the concept out in latent space against the positive prompt, a mutual cancellation rather than a filter sitting at the door.

Take those two together and a lot of folk wisdom rearranges itself.

That third row is the counterintuitive one and it matches the thing everyone has experienced at least once, where you write "no clowns" and get a clown. I don't want to overclaim the paper as an explanation for that specific failure, since the authors are describing when deletion succeeds rather than cataloguing when it backfires. But the direction is consistent.

The Vendor Telling You To Use Less Of It

Ideogram's own documentation is the most sensible thing on this SERP and it's an odd position for a vendor to take about their own feature.

Their guidance is to separate multiple items with commas, use multiple keywords to describe what you don't want, and to "Be as precise as possible without providing excessive detail". Then, more pointedly, they note that the content of the regular prompt will always be favoured over the negative prompt, and that it's "often simpler and more effective to write prompts that naturally exclude unwanted elements instead of relying heavily on negative prompts".

Their worked examples are usefully mundane. Removing green sweets from an assortment by negating "green, green color, green candies". Removing cheese from a burger with "cheese, cheeseburger, cheddar cheese".

Look at the shape of both. Three closely related terms aimed at one concept, rather than thirty terms aimed at thirty concepts. That's the pattern I'd copy, and it's roughly the opposite of the 200-item lists that rank above it.

The "write it into the positive prompt instead" advice is the part I'd underline hardest. If you don't want a person in your landscape, describing an empty valley at dawn with no path through it does more than any exclusion will, because you've given the model something to render rather than something to avoid. The general version of that principle, where describing beats instructing, is in ai image prompt.

What It Can And Cannot Do

My own table, from working with these across a few systems. The right column is where I've watched people waste time.

The quality-dial belief is worth naming directly. There's a whole genre of copy-paste negative blocks, the "blurry, low quality, watermark, deformed, bad anatomy, extra fingers, worst quality" strings, and they get pasted into every prompt forever after because the first image someone generated with them looked fine. That's not evidence. Generation is random, and the second render would also have looked fine.

I've fallen for the same thing in the positive direction, appending quality words to prompts for months because an early result was good. When I finally ran pairs with and without, I couldn't see a consistent difference. Small test, maybe a dozen pairs, so treat that as my experience rather than a result. What I'm confident about is that those blocks cost prompt space, and prompt space on images is finite in a way it isn't on text.

How To Test Whether Yours Works At All

Twenty minutes, and then you stop guessing. This is the protocol I'd run on any new system before I trusted anything I read about it.

Pick a prompt that reliably produces something you can point at. A bowl of fruit that always contains a banana works fine. Something unambiguous and countable.

Generate it four times with no negative and count how often the thing appears. Then generate four more with the negative set to that one term, nothing else. Then, and this is the step people skip, generate four more with the negative set to a nonsense term that has nothing to do with the image.

Three outcomes are possible. If the banana rate drops in batch two and not batch three, the field works. If nothing changes across all three batches, either your guidance setting has disabled it or the system doesn't really have the feature. If batch three also changed, you're seeing seed noise rather than the negative, and you need more runs before you conclude anything.

That third batch is the control and almost nobody runs one. It's the difference between knowing and believing.

Questions

How many terms should a negative prompt have? Few. Ideogram's examples use three closely related words aimed at a single concept, and that matches what I've found useful. If I'm at more than about five, I've usually got a positive prompt problem instead.

Should I keep a standard negative block? I wouldn't, not as a permanent fixture. The one exception is genuinely repetitive junk your system adds by habit, like signage and watermarks. Those are real and they recur, so excluding them is maintenance rather than superstition.

Do negative prompts work in video? Where the field exists, presumably in the same way, though I couldn't find the parameter documented on the video pages I loaded. What's worked better for me is stating what stays still, which is covered in ai video prompts.

Why does naming something in the negative sometimes summon it? The paper's neutralisation mechanism gives you a plausible story, since the concept has to be engaged in latent space in order to be cancelled. I'd stop short of saying it's proven for that failure mode.

Is there a difference between "no text" in the prompt and a negative field? Yes, and it's the difference between competing for attention inside your description and sitting outside it. Where the field exists, use the field.

What about weighting the negative terms? That's a Stable Diffusion family thing rather than a universal one, and the syntax is its own subject. It's in stable diffusion prompts.

The Habit Worth Keeping

Before you write a single exclusion, reread your positive prompt and ask whether the unwanted thing is being invited by something in it.

Nine times out of ten it is. A word like "detailed" invites clutter, "professional" invites stock photography lighting, an unset aspect ratio invites the model to compose however it likes. Fixing the invitation beats deleting the guest, and it costs you a word rather than a whole extra field.

The wider structure for image prompts is in ai image prompt, the photography-specific vocabulary sits in ai photo prompt, and the map of how all three prompt types differ is in chatgpt prompts.