Stable Diffusion Negative Prompt List (2026 Update)

A negative prompt Stable Diffusion list is the short ban string you paste into Automatic1111, Forge, ComfyUI, or a hosted SDXL box to push down watermarks, wrong mediums, and defects your checkpoint still shows. In 2026 winning lists stay short. SD 1.5 still tolerates denser cleanup. SDXL wants five to fifteen targeted terms, not a forty-line 2022 Reddit wall. You leave with paste-ready lists for photo, anime, product, and portrait jobs, a trim workflow that grows the list after you see a defect, and weight notes when a plain term fails. Soft path for a photo-driven positive: PromptMake /image. Cross-model Midjourney and FLUX negation lives on the general negative prompt guide; video bans on the video negative keywords page.

Who needs a Stable Diffusion negative prompt list

People who search negative prompt Stable Diffusion want a paste block they can drop into a dual-field UI today. They run local or hosted Stable Diffusion, not Midjourney Discord or a FLUX API that ignores negatives at default guidance. The list sits in the negative field. The positive field holds subject, light, medium, and camera. The two fields do not cancel like a checklist. Strong positive concepts still win when they fight weak negatives.

Strong fit: photographers locking documentary stills on SDXL, anime artists on Pony or Illustrious-class XL fine-tunes, product teams who need clean seamless shots without people or logos, and anyone who still maintains an SD 1.5 pipeline for a favorite checkpoint. Weak fit: Midjourney-only users who need --no, FLUX users who should rewrite the positive scene, and video creators who need temporal bans for Kling or Runway. Those stacks use different dialects. Keep this page on still Stable Diffusion.

Treat every list below as a starting paste, not a sacred spell. Checkpoint, CFG, sampler, and seed change which terms matter. Generate a baseline with an empty or near-empty negative first. Add terms for defects you saw. Delete terms that do nothing.

SD 1.5 vs SDXL: how long the list should be in 2026

Stable Diffusion versions do not share one ideal negative length. SD 1.5 bases and older fine-tunes still lean on denser quality and anatomy cleanup because the baseline model invents more mush, watermark chrome, and hand chaos. SDXL bases and mid-2026 XL fine-tunes already hold anatomy and composition better, so a bloated list flattens detail, stiffens skin, or empties backgrounds. Match list length to the family you loaded, then to the defects on the screen.

CFG still couples to negatives. On SDXL keep CFG in the 5–9 band for most photoreal and illustration work. If a new negative string makes the frame look plastic or hyper-sharpened, drop CFG by one or two points before you grow the list. On SD 1.5, higher CFG can amplify a cleanup block, but CFG 12-plus with a mega-list often cooks contrast. Tune one dial at a time: empty baseline, short list, then CFG, then weights.

SD 1.5: denser cleanup still has a job

SD 1.5 and many 1.5-era anime or realistic fine-tunes still respond to a medium-length cleanup block. A practical band is about fifteen to twenty-five terms when you group quality, artifacts, and one anatomy cluster you measured. Quality tags like worst quality, low quality, lowres, jpeg artifacts still help on datasets that used those tags in training. Artifact terms (watermark, signature, text, logo, username) remain high value. Anatomy walls help only when your checkpoint still mangles hands or faces on the poses you use.

Skip EasyNegative and other 1.5 textual inversions on SDXL graphs. They belong to the 1.5 embedding space. On 1.5 they can help as a compact quality prior if your workflow already trusts them. On XL they degrade output or do nothing. Keep embeddings matched to the base you loaded.

Starter SD 1.5 general cleanup (edit hard):

worst quality, low quality, lowres, jpeg artifacts, blurry, watermark, signature, text, logo, username, bad anatomy, bad hands, extra fingers, missing fingers, mutated hands, poorly drawn face, deformed, disfigured, cropped, out of frame

Trim anatomy terms if your fine-tune already draws clean hands. Keep artifact terms for commercial stills.

SDXL: five to fifteen targeted terms

SDXL rule for 2026: five to fifteen targeted negatives beat a cargo-cult mega-list. Lead with medium conflicts you measured (cartoon, anime, illustration for photo work; photograph, realistic, 3d render for flat illustration). Add branding artifacts (watermark, text, logo, signature, frame). Add one stubborn defect your checkpoint still shows (extra fingers, plastic skin, overprocessed). Stop. Do not paste ugly, horrible, bad composition as taste insults. Name the medium or object instead.

Starter SDXL lean baseline:

watermark, text, logo, signature, blurry

Grow from defects only. Photoreal portrait on XL often needs illustration, cartoon, anime, painting, 3d render next. Product studio often needs people, hands, clutter, oversaturated. Fantasy illustration often needs photograph, realistic, modern clothing, ui, screenshot. Architecture exteriors often need people, cars, fisheye, hdr, oversaturated sky.

Weighted negatives like (extra fingers:1.3) belong after a plain term failed on a locked seed. Start at 1.1–1.3. Past 1.5 you often crush nearby concepts. Our SDXL weights-and-negatives guide covers (term:weight) syntax and LoRA triggers in depth. This page stays on list catalogs and trim order.

Copy-paste negative prompt Stable Diffusion lists (2026)

Use these as paste cores for Automatic1111, Forge, ComfyUI negative nodes, and hosted SDXL fields that expose negative_prompt. Pick the row that matches your job. Run four samples with the positive locked. If a defect repeats, add one precise term. If a term fights the subject ("portrait" plus "no face"), fix the positive instead. Lists stay on still Stable Diffusion. Do not paste them into Midjourney as prose or into FLUX hosts that ignore the field.

Save one preset per checkpoint family, not one shared mega-file for every model on your disk. An Illustrious anime XL fine-tune and a photoreal XL base want different medium bans. Swap the medium cluster when you swap the checkpoint.

Photoreal and documentary stills

Goal: live-action photo look, natural skin or material texture, no illustration bleed, no UI chrome.

SDXL paste:

illustration, cartoon, anime, painting, 3d render, cgi, watermark, text, logo, signature, blurry, overprocessed skin

SD 1.5 paste (denser):

worst quality, low quality, lowres, jpeg artifacts, illustration, cartoon, anime, painting, 3d, watermark, signature, text, logo, blurry, bad anatomy, bad hands, extra fingers, mutated hands, deformed, oversharpened

Add plastic skin, airbrushed, hdr only after you saw that look. Add (extra fingers:1.2) on XL only after plain extra fingers failed. For portraits that need soft film grade language in the positive, keep the negative focused on medium and artifacts so you do not fight Portra-style grain you asked for.

Anime and illustration checkpoints

Goal: 2D or stylized look, block live-action photo bleed and UI screenshots. Search interest for negative prompt stable diffusion anime still clusters here.

SDXL / anime XL paste:

photograph, photorealistic, realistic, 3d render, western comic, watermark, text, logo, signature, ui, screenshot, speech bubble, bar censor

SD 1.5 anime paste:

worst quality, low quality, lowres, jpeg artifacts, photograph, realistic, 3d, watermark, signature, text, logo, poorly drawn face, bad hands, extra fingers, fused fingers, mutated hands, missing limbs, blurry

Cut anatomy terms if your anime fine-tune already draws clean hands on the poses you use. Keep photograph, realistic when the model drifts into live-action skin. Keep speech bubble, bar censor only if those show up on your set. Do not ban sketch or lineart if the positive asks for those mediums.

Product, studio, and packshot

Goal: clean product, empty seamless or controlled set, no people, no random props, no brand chrome the brief did not ask for.

SDXL paste:

people, person, hands, fingers, clutter, props, text, logo, watermark, signature, harsh shadows, oversaturated, reflection clutter

SD 1.5 paste:

worst quality, low quality, people, person, hands, clutter, text, logo, watermark, signature, blurry, jpeg artifacts, cropped, out of frame, oversaturated

If the pack includes on-pack lettering you need, remove text and logo from the negative and lock the lettering in the positive. Ideogram still wins hard typography jobs; on SD you fight lettering with exact positive strings and inpaint more than with a longer ban list.

Architecture, interiors, and empty scenes

Goal: building or room without crowd fill, cars, or sky HDR mush.

SDXL paste:

people, crowds, cars, watermark, text, logo, fisheye, hdr, oversaturated sky, cartoon, anime

SD 1.5 paste:

worst quality, low quality, people, crowds, cars, watermark, text, logo, blurry, jpeg artifacts, fisheye, deformed, oversaturated

If the brief needs staff or street life, drop people and crowds. Name the human count in the positive instead. Fisheye bans help when wide lenses invent bulge you did not ask for. They hurt when you want a deliberate ultra-wide look.

How to build and trim your own list

A copied wall fails when your checkpoint, pose, and CFG differ from the person who posted it. Build from a measured defect log. Lock seed and sampler while you test so you can see what the negative changed. Grow one cluster per retry. Delete dead weight every session. The steps below work in Automatic1111, Forge, and ComfyUI negative nodes. PromptMake /image can draft a model-aware positive from a photo and suggest SD-leaning negatives when Stable Diffusion is the target. Soft start: https://promptmake.net/image. Guests get about 3 image runs per day; a free account raises that to about 5. Edit the negative field for your checkpoint before you burn a long batch.

1. Baseline with an empty or lean negative

Run four samples with no negative, or only watermark, text on SDXL. Write the real failure: cartoon bleed, corner logo, fused fingers, crowd fill, plastic skin. If the failure is composition, crop, or identity, fix the positive first. Negatives do not invent a better camera angle or a clearer hero product.

2. Add one cluster that matches the failure

Medium conflict, artifact, object removal, or one anatomy defect. Paste three to six terms max for that cluster. Keep the rest of the settings locked. Compare to the baseline on the same seed. Keep the change only if the defect drops without killing wanted detail.

3. Weight, then escalate outside the prompt

If a plain term fails on a locked seed, try (term:1.2) once. If hands still fail, change pose, crop, seed, or inpaint before you write a novel about anatomy. If text still appears, inpaint or switch to a typography-strong workflow. CFG drops beat list growth when the frame looks overcooked after a new negative string.

Common mistakes that waste SD generations

Pasting a 2022 mega-list into every SDXL job is the top failure in 2026. XL bases punish bulk with plastic skin and empty rooms. Pasting the same wall into Midjourney or FLUX wastes the dialect. Midjourney wants --no. FLUX wants a rewritten positive scene. Pasting video flicker terms (morphing, frame flicker) into a still SD negative misses the job; those belong on video hosts.

Other traps:

• Negating the subject (portrait plus face, eyes in the negative)
• Stacking synonym spam (blurry, out of focus, soft focus, motion blur, bokeh) until the lens look dies
• Using SD 1.5 embeddings as negatives on XL graphs
• Raising CFG to 14 to "force" a weak list
• Leaving text, logo in the negative while the positive asks for packshot lettering
• Treating Reddit "best negative prompt stable diffusion" threads as checkpoint-agnostic law

Fix pattern: empty baseline → name the visible defect → add one cluster → one regenerate on a locked seed → delete terms that did nothing. Save the winner next to the checkpoint name.

When PromptMake /image helps

Use /image when a reference photo drives the positive and you want a Stable Diffusion-ready draft fast. Recreate Exactly, Change Style, Adjust Lighting, and Create Variation steer the positive. Watch any negative suggestions and trim them to defects you saw on your checkpoint. Keep SD 1.5 and SDXL presets in separate notes. Soft CTA: https://promptmake.net/image

Manual list craft still teaches the eye. Keep this page for paste catalogs and trim order. Use the cross-model negative prompt guide when you switch to Midjourney or FLUX the same week. Use the SDXL weights guide when you need (term:weight) and LoRA trigger detail.

FAQ

What is a negative prompt in Stable Diffusion?

A negative prompt is the second text field that steers sampling away from named concepts while the positive field pulls toward subject, light, and medium. You paste it into Automatic1111, Forge, ComfyUI, or any host that exposes negative_prompt. It works best with concrete nouns and medium words. Abstract insults like ugly underperform on modern XL bases.

What is the best negative prompt Stable Diffusion list in 2026?

There is no single best list for every checkpoint. SDXL wins with five to fifteen targeted terms tied to defects you measured. SD 1.5 still tolerates a denser quality and anatomy block. Start from the photo, anime, product, or architecture paste in this article, then trim. Checkpoint-specific presets beat one global mega-wall.

How long should a Stable Diffusion negative prompt list be?

On SDXL aim for five to fifteen terms. On SD 1.5 a practical band is about fifteen to twenty-five when you need quality and artifact cleanup. Longer walls from 2022 Reddit threads often flatten XL detail. Grow only after a baseline render shows a repeating defect.

Do I need different lists for SD 1.5 and SDXL?

Yes. Keep separate presets. SD 1.5 benefits more from quality tags and denser anatomy cleanup. SDXL wants medium conflicts, artifacts, and one stubborn defect. Do not port EasyNegative or other 1.5 embeddings into XL graphs. Match embeddings and list density to the base you loaded.

Where do I put a negative prompt Stable Diffusion string?

Put it in the dedicated negative field, not at the end of the positive prompt as a sentence. ComfyUI users wire a negative conditioning input on the text encode or sampler path their graph uses. Hosted APIs send it as negative_prompt. Midjourney and many FLUX hosts do not use this field the SD way.

Should I use weighted negatives like (extra fingers:1.3)?

Use weights after a plain term failed on a locked seed. Start near 1.1–1.3. High weights crush nearby concepts and can stiffen the whole frame. Fix pose, crop, seed, or inpaint before you stack five weighted anatomy terms. The SDXL weights guide on this blog covers syntax in more depth.

How do I start for free with a photo reference?

PromptMake /image turns a photo into a draft aimed at Stable Diffusion and other image targets. Guests get about 3 image generations per day; registered free accounts get about 5. Paste the positive into your SD UI, then apply a short list from this page. Start at https://promptmake.net/image