Negative Prompt Guide

1. What a negative prompt actually does

A negative prompt is the companion text that tells the model what not to include. Under the hood, it's implemented by classifier-free guidance (CFG): the model computes two predictions — one with the positive prompt, one with the negative — and steers away from the negative. That means your negative prompt isn't just a wishlist; it actively shapes the final sampling direction on every step.

Fun fact: Many experienced users report that a good negative prompt has more impact on final quality than doubling the positive token count.

2. The basic set — universal baseline

Start every project with this set. It covers the three most common failure modes: low sample quality, anatomical errors, and unwanted overlays (watermarks, text).

low quality, worst quality, normal quality, poorly drawn, blurry, lowres, jpeg artifacts, watermark, signature, text, error, logo, username, artist name

bad anatomy, bad hands, missing fingers, extra fingers, fused fingers, extra limbs, missing limbs, deformed, disfigured, mutation, mutated, cross-eyed, bad proportions

Sizing: Keep the basic set under 20 tokens. Anything more starts to restrict creativity — the model literally has fewer directions it can go.

noneSafety guard only

guard only+ anatomy terms

quality terms+ quality terms

full stackFull stack

Shared prompt

Differs per cut — each cut on the comparison page has it in full.

Model

rinFlanimeIllustrious_v30 · euler_ancestral/normal · steps 28 · CFG 5 · 896×1152 · seed 777004

This comparison was measured on a single Stable Diffusion family checkpoint. The same tags can behave differently on NovelAI, Midjourney or another checkpoint — read it as a reference point, not a verdict.

3. The advanced set — when the basic isn't enough

Add these when you consistently see specific problems. Don't stack everything at once — add one group at a time.

asymmetric eyes, cross-eyed, wonky eyes, misaligned eyes, ugly face, bad face, deformed face, distorted face, plastic face

bad hands, mutated hands, poorly drawn hands, extra fingers, missing fingers, bad feet, extra toes, fused toes, long fingers, short arms, long arms

out of frame, cropped, cut off, off-center, bad framing, poorly cropped, head out of frame

duplicate, clones, tiling, repeating pattern, multiple heads, twin body, fused bodies

oversaturated, overexposed, underexposed, harsh shadows, motion blur (unintended), grainy, noisy, purple fringing

4. Five situational templates

Copy these directly. Each template combines the basic set with the specific hazards of that scene type.

low quality, worst quality, blurry, jpeg artifacts, watermark, text, bad anatomy, bad hands, missing fingers, extra fingers, cross-eyed, asymmetric eyes, plastic face, ugly, deformed, out of frame, cropped

low quality, worst quality, blurry, oversaturated, cartoon, anime, watermark, text, logo, people, person, human, buildings (unless intended), duplicate, tiling, deformed terrain

low quality, worst quality, normal quality, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra limbs, ugly, deformed, text, watermark, signature, artist name, username, realistic (if you want anime), photo, 3d

low quality, worst quality, blurry, jpeg artifacts, watermark, text, cartoon, anime, illustration, painting, 3d render, plastic skin, waxy skin, airbrushed, deformed, bad anatomy, bad hands

low quality, worst quality, blurry, watermark, text, daytime, clear sky, sunny, rural, nature (unless intended), medieval, fantasy armor, boring composition, flat lighting

5. Model-specific caveats

• SDXL / A1111: All of the above works directly. Put the full negative prompt in the negative field.
• NovelAI: Use the UC (Undesired Content) field. Their preset "Heavy" option covers most of the universal basic set.
• Animagine XL / Pony V6: Include lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry (recommended preset).
• Flux.1-dev / schnell: Negative prompts are not supported. Use positive phrasing instead ("clean skin" rather than negative "blemishes").
• Midjourney: Use --no flag. --no text, watermark, blurry, low quality is the common shorthand.

6. Attribute leakage — when negatives backfire

A negative prompt can leak into the positive and remove things you wanted:

• Color leak: If you want red hair, do not put red in the negative — the model may remove red altogether. Be specific: red dress only.
• Subject leak: Putting person in the negative kills your character. Be specific: crowd, background people.
• Style leak: painting in the negative will remove all painterly qualities — fine for photos, bad for illustrations.

7. Iteration workflow

1. Start with the basic set + one situational template.
2. Generate 4 images, identify the worst common problem.
3. Add 1–2 specific negative tokens targeting that problem.
4. Regenerate with the same seed. If the problem persists, the issue is probably in the positive prompt, not the negative.
5. Save your winning combo in Promgrammer for reuse.

8. Related

For more context on when and why these tokens matter: Quality Tags Deep Dive, Model Comparison.