Negative Prompts for AI Video: What to Exclude (and Why It Works)

Most AI video tutorials teach you what to put in a prompt. Almost none teach you what to leave out. That’s a problem because half the difference between a usable AI video clip and an unusable one is what you’re telling the model to avoid generating, not what you’re telling it to generate.

Negative prompts are how you do that. Some video models support them as a dedicated input field; others require you to fold them into your main prompt. Either way, getting them right is the fastest way to escape the loop of regenerating the same broken clip with slight wording changes.

Here’s how negative prompts actually work across the major AI video tools, what to put in them, and what’s a waste of time.

What a negative prompt actually is

A negative prompt is a list of things you don’t want in your generated video — visual artifacts, unwanted styles, specific objects, certain camera behaviors, or technical defects.

In tools that support a dedicated negative prompt field (Kling, Hailuo, some Runway modes), you type these in a separate box. The model treats them as steering signals away from those features during generation.

In tools that don’t have a separate field (Sora, Veo, most Runway modes), you append the exclusions to your main prompt with phrasing like “no blur, no warped faces, no extra fingers, no lens flare.” This is less precise than a dedicated field — the model may still treat the words “lens flare” as a positive feature in some cases — but it works often enough to be worth doing.

The mental model: positive prompts pull the output toward what you describe. Negative prompts push the output away from what you describe. They work together, not in opposition.

Why negative prompts matter for AI video specifically

Image generators have used negative prompts for years. AI video adds a few new failure modes that make negative prompts more important:

Frame-to-frame inconsistency. A person’s face can subtly morph between frames. “Consistent face, no morphing” in your negative prompt steers against this.
Hand and finger artifacts. Hands are still the failure mode for most video models. Excluding “extra fingers, fused fingers, deformed hands” is almost always worth doing if hands are visible.
Camera shake and motion blur. Some models default to over-stylized motion. Negative-prompting “shaky cam, motion blur, blurry” cleans this up.
Unwanted text generation. Models sometimes hallucinate signs, watermarks, or text overlays. “No text, no watermark, no logo” stops this.
Style drift. Asking for a “natural lighting documentary” can produce something that drifts toward cinematic with heavy color grading. Negative-prompting “stylized, oversaturated, cinematic LUT” keeps it grounded.

These problems aren’t always solved by better positive prompts because they’re defaults the model falls into when uncertain. Negative prompts give the model a signal about what to avoid when it doesn’t have a strong instruction for what to do.

Tool-by-tool: how negative prompts work

Sora (and Sora 2)

Sora doesn’t have a dedicated negative prompt field. Append exclusions to your prompt:

A close-up of a chef chopping vegetables in a sunlit kitchen.
Natural color grading, steady camera. Shot on 50mm lens.
Avoid: motion blur, over-saturation, deformed hands, extra fingers,
text overlays, watermarks, frame artifacts.

Sora responds better to “Avoid:” or “No:” sections than to embedded negation in a sentence. Putting them at the end is more reliable than scattered through the prompt.

Veo 3

Same pattern as Sora — no dedicated field, append at the end. Veo responds particularly well to negative-prompting style elements:

[Positive prompt here]
Avoid: cartoonish style, over-stylized lighting, artificial bokeh,
shaky camera, low resolution, washed out colors.

If you’ve read our Veo 3 prompt guide, the same principles apply — Veo wants explicit, declarative language. That goes for negation too: “Avoid: warped faces” works better than “without facial distortion.”

Runway Gen-4

Runway has a dedicated negative prompt field in some modes (specifically the newer Gen-4 image-to-video). When available, use the field rather than appending. Runway’s negative prompt is weighted more heavily there than text-embedded exclusions.

Negative prompt field:
distorted face, deformed hands, extra fingers, blurry, low quality,
oversaturated, motion blur, shaky cam, watermark, text, low resolution

For text-to-video modes without the field, append the exclusions to your main prompt.

Kling

Kling has a dedicated negative prompt field that’s heavily weighted. Use it. Kling’s positive prompts can drift toward over-stylization, so a strong negative prompt is essential for natural-looking output:

Negative prompt:
3D render, CGI, cartoon, anime, animated, plastic skin, doll-like,
deformed, distorted, blurry, watermark, text overlay, motion blur,
warped face, extra limbs, fused fingers

Kling responds dramatically to “anime” or “cartoon” in the negative prompt — even non-anime subjects can come out with a stylized look without those exclusions.

Hailuo (MiniMax)

Hailuo also has a dedicated negative field. Treat it similarly to Kling. Hailuo specifically benefits from negative-prompting “low contrast” and “muddy colors” — it tends toward those when uncertain.

Pika

Pika has limited negative prompt support depending on which model version you’re on. The newer Pika supports a negative input; older versions don’t. When in doubt, append.

A template negative prompt that works for most realistic video

There’s no single perfect negative prompt, but for realistic, natural-looking video output across most models, this template is a strong default:

deformed, distorted, blurry, low quality, low resolution,
extra fingers, fused fingers, warped face, plastic skin,
motion blur, shaky camera, watermark, text overlay, logo,
oversaturated, washed out, cartoonish, 3D render, CGI

Adjust by removing items that conflict with what you actually want. If you’re going for a 3D-rendered look, obviously don’t negative-prompt “3D render” or “CGI.” If you want stylized cartoon output, drop “cartoon” from the list.

A template negative prompt for stylized / animated video

For anime, cartoon, or stylized output, the inverse:

photo, photorealistic, realistic, plastic, doll-like,
deformed, distorted, blurry, low quality, watermark, text,
extra limbs, warped proportions, motion blur, shaky camera

Notice “photo” and “photorealistic” in the negative for stylized output — without those exclusions, models often try to compromise toward photorealism and produce uncanny middle-ground results.

What to put in negative prompts (categorized)

Quality issues (almost always include):

blurry, low resolution, low quality, pixelated, compression artifacts

Anatomy issues (include when humans are visible):

deformed hands, extra fingers, fused fingers, missing fingers, warped face, asymmetric eyes, distorted body, extra limbs

Camera issues (include unless intentional):

motion blur, shaky camera, lens flare, chromatic aberration, vignette

Style issues (depends on intent):

oversaturated, washed out, cartoonish, photorealistic, anime, 3D render, CGI, plastic skin

Frame issues (almost always include):

watermark, text overlay, logo, subtitle, caption, signature, frame border

Inconsistency issues (include for longer clips):

morphing face, character drift, inconsistent lighting, frame jumps

What NOT to put in negative prompts

Things that don’t actually work and just clutter the prompt:

Specific brand names (“no Coca-Cola logo”) — the model isn’t trained to recognize and avoid specific brands; it’s better to specify a generic alternative in the positive prompt
Plot or narrative concepts (“no sad ending,” “no violence”) — negative prompts steer visual features, not narrative structure
Camera-specific exclusions in detail (“no Sony A7S III color profile”) — too specific; the model doesn’t have that resolution of camera knowledge
Negation of features you didn’t ask for (“no underwater scene” when you asked for a kitchen) — wastes tokens; the model wasn’t going there anyway
Vague concepts (“no bad video,” “no ugly,” “no weird”) — too abstract; specify the concrete artifact you want to avoid
Excessive length — a negative prompt of 50 items is worse than one of 15 well-chosen items. The signal gets diluted.

How to debug a bad output with negative prompts

When you get a bad output, identify the specific failure mode and add it to the negative prompt before regenerating:

Bad hands? Add “deformed hands, extra fingers, fused fingers”
Face morphing between frames? Add “morphing face, inconsistent face, character drift”
Output looks plastic / fake? Add “plastic skin, doll-like, 3D render, CGI”
Camera too shaky? Add “shaky camera, handheld, motion blur”
Random text appearing? Add “watermark, text overlay, subtitle, signature”
Style drifting away from realistic? Add “stylized, cartoonish, oversaturated”

Each regeneration, narrow the negative prompt based on what you actually saw. This is faster than rewriting the positive prompt from scratch.

Negative prompts vs prompt rewrites

Sometimes the right answer is a stronger positive prompt, not a longer negative prompt. The rule of thumb:

If the model is doing something you didn’t ask for, use a negative prompt to push against it
If the model is failing to do what you asked for, rewrite the positive prompt with more specificity

For example: if you asked for a “calm beach scene” and got dramatic crashing waves, that’s a positive prompt issue — specify “still water, gentle ripples, no waves” in the positive prompt. If you asked for a beach scene and got a beach scene with a random watermark, that’s a negative prompt issue — exclude “watermark.”

Knowing the difference saves regeneration time. (For positive prompt structure, see our AI video prompt structure formula guide.)

Quick reference: negative prompt by use case

Product video (e-commerce):

deformed, blurry, low quality, watermark, text overlay,
motion blur, shaky camera, oversaturated, plastic look,
extra parts, distorted product, lens flare

Talking-head / interview style:

warped face, morphing features, extra teeth, deformed mouth,
inconsistent face, plastic skin, low resolution, blurry,
heavy motion blur, shaky cam, watermark, text

Cinematic narrative shot:

amateur footage, low quality, washed out, oversaturated,
cartoonish, anime, 3D render, watermark, text overlay,
deformed, motion blur, shaky camera, vignette, lens flare

Real estate / architectural:

distorted perspective, warped lines, blurry, low resolution,
oversaturated, cartoonish, watermark, text, motion blur,
people, vehicles, deformed structure

Animated / stylized:

photo, photorealistic, realistic, plastic, deformed,
warped proportions, blurry, low quality, watermark, text,
extra limbs, motion blur

FAQ

Do all AI video models support negative prompts?

Most modern AI video tools support negative prompts in some form. Kling, Hailuo, and newer Runway modes have dedicated negative prompt fields. Sora, Veo, and older Runway modes don’t have a separate field — you append exclusions to your positive prompt instead. The functionality is similar; the implementation varies.

How long should my negative prompt be?

10 to 20 well-chosen terms is the sweet spot. Beyond that, signals get diluted and the model may start ignoring some entries. A focused list of the specific artifacts you want to avoid beats a long generic list.

Can I use the same negative prompt for every video?

A baseline negative prompt covering quality issues, anatomy issues, and watermarks works for most realistic video. Save it as a template and append a few use-case-specific exclusions per project. For stylized or animated content, you’ll need a different baseline.

Why does my video still have the artifacts I negative-prompted?

A few possibilities: the negative prompt was too vague (e.g., “no weird stuff” doesn’t work — you need specific artifacts), the model doesn’t strongly support negative prompts (older Pika versions, for example), or the artifact is so strongly weighted in the positive prompt that the negative can’t overcome it. Try shortening the positive prompt or moving the description to a different framing.

Should I include negative prompts when using image-to-video?

Yes — image-to-video can introduce its own artifacts during the animation step. Include exclusions for “morphing face,” “character drift,” and “frame inconsistency” specifically. The starting image is fixed, but the motion can introduce instability the negative prompt helps control.

Do negative prompts affect generation time or cost?

Generally no. Negative prompts add a small amount of processing but don’t significantly increase generation time or token cost on any major platform. They’re cheap to use, so use them.

Wrapping up

Negative prompts are the unglamorous half of AI video prompting. Most tutorials skip them because positive prompts are more interesting to write about. But if you’re spending hours regenerating the same clip trying to fix the same artifacts, the missing piece is almost always a stronger negative prompt.

Build a template list, customize per project, and add new exclusions whenever you see a new artifact. The compound effect on output quality is significant.

If you want a head start, LzyPrompt generates structured prompts with negative-prompt sections built in — pick your tool, describe your scene, and get both halves of the prompt ready to paste.