Negative Prompts Explained: Examples for Images and Video

A negative prompt is a short list of things you want the model to steer away from: blurry detail, extra fingers, watermarks, a style you never asked for. It works as a push in the opposite direction from those words, not as an eraser, so it cleans up recurring flaws but can't rescue a vague main prompt. Used lightly, it is one of the cheapest ways to get cleaner images and steadier video.

What a negative prompt actually does

Most current image and video models build a result over many steps, leaning toward what your prompt describes at each one. A negative prompt gives them a second target to lean away from. Your main prompt says "go here", the negative says "and not there".

That's why concrete visual terms work so well. "Watermark" and "extra fingers" are things the model recognizes on sight. Vague terms like "ugly" give it almost nothing to push against.

What it can't do

It doesn't work like a delete key. If your main prompt describes a busy hotel bar and you put "people" in the negative, you are arguing with yourself, and you'll get a muddled compromise. Change the main prompt instead.

It also doesn't understand "no" in the main prompt. Typing "no hat" there often produces a hat, because the model latches onto the noun. Move the word to the negative field and write it plainly: "hat", not "no hat".

And it won't fix a bad setup. A pose description that crams four actions into one sentence will keep producing broken anatomy no matter how many hand terms you list. Simplify the prompt first.

Where it lives on SpicyGen

On SpicyGen the negative prompt sits under Advanced when you generate images, animate an image, or describe a video from text. The image field takes up to 600 characters; if you hit that limit, your list is too long. For text to video, a blank field uses defaults, so only write one when something specific keeps going wrong.

The art style (photorealistic, studio, anime, 3D or cartoon) is its own setting, and lighting, camera angle and mood go in the main prompt. The negative is for exclusions, not direction. You'll find it in the adult AI image generator.

How to build one that works

Start with nothing. Generate once with your main prompt, look hard at what is actually wrong, and add terms for those failures only. Then run the same prompt again so you can tell whether the change helped.

A typical round on a realistic portrait: the first result looks airbrushed, so "plastic skin, airbrushed" goes in. The second has a mangled hand, so "extra fingers, fused fingers" goes in. The third has faint text in a corner, so "text, watermark" goes in. Six terms, each one earned:

Prompt: a woman in her thirties with shoulder-length auburn hair, sitting on a velvet sofa in a dim hotel suite, warm lamp light, 35mm photo, shallow depth of field

Negative: plastic skin, airbrushed, extra fingers, fused fingers, text, watermark

Some habits that save time:

• Write short comma-separated phrases, not sentences.
• Pick the specific failure over the general one. "Fused fingers" beats "bad anatomy".
• Never negate something the main prompt asks for.
• Use positive phrasing when it exists. Write "bright midday sun" in the prompt rather than putting "dark" in the negative, which can also wipe out shadows you liked.

Realistic, anime and 3D need different negatives

The style you pick decides half your negative prompt, because one of the most useful things to exclude is every other style.

Realistic

Photorealistic results tend to fail in two directions. They drift toward illustration, or the skin goes waxy like an over-retouched ad. So negate the other styles (cartoon, anime, 3D render, painting) and the polished look (plastic skin, airbrushed), then add hand terms. We would leave "grainy" out unless grain is a real problem: a bit of texture is a big part of what makes photorealistic adult images read as photos rather than renders.

Anime

Flip it. "Photorealistic", "photo" and "3D render" go into the negative, since anime output sometimes slides toward semi-real shading. The other common flaws are muddy colors, messy or doubled lines and mismatched eyes. Never paste a realistic negative here: "anime" and "illustration" in it will fight the style you selected. The same reasoning carries over when you make anime videos.

3D

3D sits in between. Exclude flat looks on one side (sketch, flat shading, anime) and photographs on the other, then add render problems like blurry textures and clipping. "Plastic" is a judgment call: a stylized 3D character is supposed to look a little like a figure, so add it only if the skin reads cheap.

Negative prompts for video

Video adds a new way to fail: time. Every frame can look fine on its own while the clip flickers, the face slowly morphs into someone else, a third arm appears during a turn, or nothing moves at all. Those are the words a video negative needs: flicker, jitter, warping, morphing face, extra limbs, static, frozen frame.

Watch for terms that pull against each other. "Static" asks for more motion, while "camera shake" and "jitter" ask for less. Both can sit in one negative, but if a clip comes out hectic, drop "static" first.

Image to video is a different job

When you animate an image from your library, the starting frame has already settled the look, so style terms like "cartoon" or "3D render" do very little. Spend the negative on motion problems and on things that appear from nowhere mid-clip, like captions or a second person. The bigger lever is the motion prompt itself, which our guide to writing image to video prompts covers.

Why long negative prompts backfire

Forums are full of negative prompts that run to several paragraphs, and copying one feels like insurance. It usually makes results worse.

Each term gets less influence as the list grows, so the two flaws you care about are diluted by forty you never had. Long lists also pick up terms that collide with your main prompt: "dark" in a moody bedroom scene, "blurry" when you asked for shallow depth of field. Some text encoders only read so far, so the end of a giant list may barely register. Worst of all, a heavy negative flattens everything. Push away from enough things and the model retreats to a safe middle where every image has the same glossy, evenly lit look.

Our rule of thumb is 8 to 15 terms. If a term hasn't fixed something you've actually seen, take it out.

Keeping characters clearly adult, and keeping extras out

Age doesn't belong in the negative field. Adulthood is something you describe, not something you subtract, so put it in the main prompt in plain words: "a woman in her thirties", "a man in his forties". Back it up with adult context like the setting and wardrobe. Once a character looks right, save it as a character so the same adult face carries across every image and video instead of drifting.

That matches how SpicyGen works on its side. Every prompt is screened before anything is generated, and every image and video gets an AI age check before you see it. Anything suggesting someone under 18 is blocked, along with the other categories in the user guide.

Unwanted extras, on the other hand, are exactly what negatives are good at. Captions, watermarks and logos turn up uninvited all the time. So do extra people: if a second figure keeps wandering into a one-person scene, add "extra person, second figure, crowd".

Starter negative prompts to copy

Paste one, generate, then trim anything that isn't earning its place.

Realistic photo

cartoon, anime, 3d render, illustration, painting, plastic skin, airbrushed, oversaturated, blurry, extra fingers, fused fingers, deformed hands, text, watermark, logo

Anime

photorealistic, photo, 3d render, muddy colors, messy lines, mismatched eyes, blurry, extra fingers, deformed hands, text, watermark, signature

3D

photo, anime, sketch, flat shading, low poly, blurry textures, clipping, extra fingers, deformed hands, text, watermark

Text to video

flicker, jitter, warping, morphing face, face changing, extra limbs, distorted hands, static, frozen frame, blurry, text, subtitles, watermark

Image to video

warping, flicker, morphing, face changing, extra limbs, sudden cut, camera shake, text, subtitles, watermark

Add-on for one-person scenes

extra person, second figure, crowd, duplicate body

One habit worth picking up: when a negative fixes a specific problem, save it next to the prompt that needed it. A saved character keeps the face consistent, and reusing the same negative with that character keeps the finish consistent too, so a set of ten images looks like one shoot instead of ten experiments.