Start once. Return to a new image every morning.

Negative Keywords Guide

Negative keywords tell the model what to avoid. They help when your main prompt keeps pulling in the same mistake. You do not need them for every stream. Use them when a pattern keeps showing up and you want to push the model away from it.

Why They Help

Your main prompt tells the model what to draw. Negative keywords trim away the junk that keeps leaking in. That can mean anatomy errors, stray text, watermarks, duplicate subjects, bad crops, or the wrong visual style.

Use them when you notice:

  • text or logos appearing when you did not ask for them
  • extra limbs, fingers, or faces
  • messy crops around hands, feet, or the top of the head
  • the model drifting into anime, CGI, or painterly output when you want realism
  • NSFW details creeping into a prompt that should stay clean

How To Write Them

Keep the list short and blunt. Start with the specific problem you are trying to stop. Add more only if the same fault keeps coming back.

Good pattern:

watermark, text overlay, extra fingers, cropped hands, duplicate subject

Bad pattern:

bad image, ugly, do not make anything weird or low quality or strange

Practical rules:

  • Use comma-separated phrases.
  • Name visible faults, not abstract complaints.
  • Do not dump fifty terms in on day one.
  • If the model starts ignoring your main prompt, cut the list back.

Starter Lists

Portraits

Use this when faces, hands, or body proportions keep slipping.

extra fingers, extra hands, malformed hands, deformed face, asymmetrical eyes, duplicate face, fused fingers, bad anatomy, cropped head, cropped hands, waxy skin, blurry face

Products

Use this when you need clean packshots, product renders, or editorial still life.

watermark, logo, text overlay, distorted packaging, duplicate product, warped edges, broken reflections, messy background, clutter, cropped object, low detail

Food

Use this when ingredients melt together or the plate looks synthetic.

plastic texture, duplicate food, extra utensils, floating garnish, messy plating, unnatural colours, warped plate, blurry detail, text, watermark

Interiors And Architecture

Use this when rooms bend, furniture duplicates, or windows go strange.

warped walls, bent lines, duplicate furniture, floating objects, distorted perspective, clutter, low detail, text, watermark, broken windows

Landscapes And Nature

Use this when the scene gets muddy or artificial.

oversaturated sky, duplicate trees, blurry foreground, artificial lighting, low detail, text, watermark, distorted horizon, muddy texture

Illustration And Anime

Use this when characters break down or you want cleaner composition.

extra limbs, extra fingers, duplicate character, off-centre eyes, bad anatomy, cropped hands, cropped feet, text, watermark, messy background

Typography And Signage

Use this when the scene should contain signs, labels, or printed layouts but the model keeps smearing them.

illegible text, gibberish lettering, duplicate letters, warped sign, extra signage, watermark, logo, blurry print, cluttered layout

NSFW Suppression

Negative keywords can reduce NSFW drift. They do not replace moderation. If a prompt sits close to the line, the safest fix is still to rewrite the main prompt so the subject stays unambiguous.

Starter list for keeping a stream clean:

nudity, nipples, areola, exposed breasts, genitalia, explicit, erotic, sexual act, lingerie, fetish, porn, nsfw, topless, see-through clothing

Use it when:

  • the model keeps adding exposed skin you did not ask for
  • swimwear, underwear, or glamour prompts keep drifting into explicit territory
  • you want a safer baseline before sharing a stream publicly

Start Small

If you are not sure where to begin, start with one of these:

  • Realistic portrait: extra fingers, malformed hands, blurry face, cropped head, watermark
  • Product shot: watermark, text overlay, duplicate product, clutter
  • Safe general-purpose list: watermark, text overlay, duplicate subject, extra limbs, nudity

Then watch the next image. If the mistake stops, leave the list alone. If the mistake changes shape, replace the weak term with the new problem instead of piling on more noise.