Y
Prompt Workshop
Guide · Negative prompts

How to Write Effective Negative Prompts in Stable Diffusion

The goal of a negative prompt is to block the failures your prompt is actually likely to produce, not to enumerate every imperfection you have ever seen. This guide gives short, targeted templates for five high-frequency scenarios and explains the differences between SD 1.5, SDXL, Flux and Midjourney.

Why longer negative prompts make things worse

Beginners copy a wall-of-text negative prompt and assume more is better. Negative prompts also consume model attention. Listing "asymmetric eyes, cross-eyed, deformed pupils, wrong eye color, extra eyes" five times does not make eyes more accurate — it makes the eye region weighted up, sometimes producing odd makeup, oversaturated irises or unnatural highlights.

A more reliable approach: pick the 3–5 problems your current prompt is genuinely prone to and ignore everything else. A side-profile prompt does not need asymmetric eyes. A silhouette in a wide shot does not need bad hands. Targeting beats coverage.

SDXL and Flux reduce the need for negative prompts dramatically compared to SD 1.5. Flux Dev and Schnell, built on a Transformer architecture, essentially ignore negative prompts; you control quality through positive prompting alone. Knowing your model generation is the prerequisite for everything that follows.

5 scenario templates

ScenarioNegative prompt template (copy-paste)
Realistic portraitdeformed face, asymmetric eyes, plastic skin, over-smoothed, low-res, blurry, watermark, signature
Hand close-upextra fingers, fused fingers, missing fingers, malformed hand, wrong proportions, blurry hand
Poster with textmisspelled text, broken letters, gibberish, random characters, distorted font, low-res text
Low-res / repairlow-res, jpeg artifacts, blurry, noisy, oversharpened, compression artifact, banding
Video stabilitycamera shake, flicker, jitter, frame skipping, ghosting, motion blur on static subject

Each template caps at eight tokens. Need both portraits and text in the same image? Merge the two, but stay under 12 tokens total.

Weight syntax: (token:1.3) and [token:0.7]

(deformed hand:1.4)strongest suppression — use only when hands repeatedly fail
(blurry:1.2)mid-weight suppression for sharpness issues
watermarkdefault 1.0 — no need to amplify
[oversharpened:0.6]weak negative — avoids over-correction

Pushing weights above 1.5 backfires: the model may avoid the concept entirely (hide the hand behind a sleeve) rather than fix it. Keep negative weights between 0.6 and 1.4.

Wrong vs. right examples

✗ Wrong (mega list)

worst quality, low quality, normal quality, lowres, monochrome, grayscale, watermark, signature, ugly, deformed, mutated, mutation, bad anatomy, bad proportions, gross proportions, text, error, missing fingers, extra digits, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry, bad feet, cloned face, fused fingers, too many fingers, long neck, malformed limbs, asymmetric eyes

30+ tokens, duplicated entries (watermark, signature appear twice). Result: worse than the 6-token targeted template below.

✓ Right (scenario-targeted)

(deformed face:1.3), asymmetric eyes, plastic skin, blurry, watermark, low-res

Targets the five failures most likely on realistic portraits. Short, no duplicates, weights only on the highest priority.

SDXL vs. Flux: a generational shift

SD 1.5 made negative prompts essential — every image used a long list. SDXL halves the dependency. Flux Dev and Schnell, built differently, treat negative prompts as low-priority noise. On Flux, the better move is to phrase the requirement positively: "clean composition, single subject, no extra elements" reads more strongly than a negative list.

Practical guidance: SD 1.5 uses 8–12 negative tokens; SDXL uses 4–8; Flux Dev typically uses none, or phrases the requirement positively; Midjourney uses --no with one to three concrete tokens.

4 common mistakes

Mistake 1 · Reusing the mega list across model generations

That wall-of-text template was tuned for SD 1.5. On SDXL or Flux it drags the result down.

Mistake 2 · Positive words inside the negative list

"Low quality, bad quality" in negative + "high quality, masterpiece" in positive cancel each other. Send one signal per concept.

Mistake 3 · Negative tokens that do not apply to your scene

White-background product shots do not need NSFW negatives. Each negative token should target a failure the prompt is actually prone to.

Mistake 4 · Using negatives to fix vague positives

A vague "a woman" cannot be rescued by "not ugly, not old". Specify the subject first; negatives are insurance, not a corrective layer.

Frequently asked questions

Does Midjourney support negative prompts?

Yes, via the --no parameter. Use it with a comma-separated list of concrete tokens (--no people, text, watermark), not full sentences.

Does Flux really ignore negative prompts?

Flux's architecture barely uses them. Writing positive constraints ("single subject, clean white background") is the modern equivalent.

Should LoRAs change my negatives?

Sometimes. If a LoRA learned a tendency (e.g. oily skin on a realistic-portrait LoRA), add a targeted negative like (oily skin:1.2).

What weight should I use?

Stay between 0.6 and 1.4. Above 1.5 the model often avoids the concept entirely instead of fixing it.

Try this skeleton in the structured editor

Open the editor and fill in subject / style / light / composition blocks separately; the editor assembles the final prompt for you.

Open the editor →
Yan · AI Prompt Workshop editorial team|Last updated on 2026-06-12。This site does not call any cloud model. Every prompt and parameter in this article was tested and refined locally by the editorial team.