← 提示词库 在 GitHub 查看源文件 ↗
中文

Negative Prompts — Reusable Anti-Artifact Block

Authored by jnMetaCode (MIT). A companion to the atmosphere prefabs: where those describe what you want, this file collects what you want to exclude — the recurring failure modes of AI video (melting hands, plastic skin, frame flicker, the "video-game look") and how to suppress them per model.

How to use: drop the canonical block below into your model's negative field if it has one; otherwise read the compatibility table — several tools need positive-only phrasing instead, and a couple will summon the very thing you list if you write "no ___".

中文版 →


The canonical negative prefab

Copy-paste this as your default. It targets resolution/sharpness defects, overlay junk, anatomy breakage, the over-stylized CG/anime look, and dead camera work — the artifacts that read as "AI" at a glance.

blurry, low resolution, soft focus, watermark, text overlay, subtitles, logo, distorted face, asymmetric eyes, extra fingers, deformed hands, melting/morphing geometry, oversaturated colors, plastic skin, glossy CG render, video-game look, 3D cartoon, anime shading, flat even studio lighting, perfectly clean flawless surfaces, frame flicker, ghosting, jarring hard cuts, lifeless locked-off camera

It is written as a plain comma list of unwanted nouns/phrases on purpose — that is the format the dedicated negative fields (Veo, Kling, Wan, Hailuo) expect. Do not prefix items with "no" or "don't" when you paste into a dedicated field. For tools without a field, see the in-prompt negation guidance below.


Which models take a negative field — and how to feed it

Mechanics differ more than people expect. Some tools have a dedicated box, some need the negatives folded into the positive prompt as guardrail phrases, and two (Runway Gen-4, and to a degree Seedance's consumer app) do not reliably support negation at all — on Runway, "no X" can actively summon X.

ModelDedicated negative field?How to negateSingle-shot maxIP filterLang
Seedance 2.0 (Doubao / Jimeng-Xiaoyunque, ByteDance)Partial — no reliable field in the consumer appDescribe what you DO want; fold avoidance phrasing in-prompt~10–15s (Doubao app locked to 5s/10s buttons; Jimeng web & VolcEngine console expose 4–15s)Strict (domestic ByteDance is harsher than most Western tools)Either (native ZH safest)
Veo 3 / 3.1 (Google)YesPlain noun/phrase list in the field — no "no/don't" commands8s (4/6/8s); Extend stacks 7s hops up to ~148s, quality drops after 4–5StrictEN
Kling 2.x / 3.0 (Kuaishou)YesBest for stability artifacts (sliding feet, extra fingers, morphing), not generic "quality" words2.5: 5–10s (Pro ~12s); 3.0: ~15s single-promptStrict (banned-word filter rejects the WHOLE prompt on one match)Either
Hailuo / MiniMax (Hailuo 02 / 2.3)YesUse sparingly — boundary-setting for specific artifacts, not a primary lever~6–10s (1080p ~6s; 768p ~10s)ModerateEither
Wan 2.x (Alibaba, open-source)Yes (robust)Documented defaults: morphing, warping, face deformation, flickering, jittering, inconsistent lighting2.2: ~3–8s; 2.5/2.6: ~10–15sLenient when self-hostedNative ZH (EN-only can underperform)
Runway Gen-4 / 4.5NoPositive-only — describe only what SHOULD appear; "no X" can produce the opposite5s or 10sStrictEN
Pika (2.2 / 2.5)Partial — yes on 2.5, unclear on 2.22.5 accepts "no morphing, no extra limbs…"; verify in-app on 2.25s/10s; Pikaframes keyframes up to ~25sModerateEN
Sora 2 / 2 Pro (OpenAI)No dedicated boxNegate in-prompt with guardrail phrases ("original characters only, no logos or text")Up to ~25s on Pro (standard tier shorter)Strict (triple-layer; catches lookalikes, not just names)EN

Three rules that trip people up:

  1. Runway Gen-4 is positive-only. Porting no distorted hands into a Runway prompt is the single most common cross-tool mistake — it can summon distorted hands. Rephrase the whole intent positively ("clean five-fingered hands, anatomically correct").
  2. Dedicated fields take descriptions, not commands. In Veo/Kling/Wan, put extra limbs, glitch morphs, text overlays — never no extra limbs. The "no" wastes a token and can confuse the parser.
  3. Seedance's consumer Doubao app has no surfaced negative field and is locked to 5s/10s buttons. Don't promise 15s or a negative box there; the full 4–15s range and finer control live on Jimeng web / VolcEngine.

Confidence: durations and negative-prompt mechanics are medium-high for Veo 3.1, Runway Gen-4, Kling, Wan, and Sora (consistent across 2026 vendor/help docs); medium for Seedance 2.0 and Hailuo (mostly third-party guides). Veo's ~148s and Sora/Pika's ~25s come from extension/keyframe features, not plain single-shot generation. IP-filter "strictness" labels are qualitative judgments, not formal vendor classifications.


Scenario-tuned variants

Append the relevant line to the canonical block for the genre you're shooting. These target failure modes specific to each scenario rather than generic quality.

Transformation sequences

premature full transformation, symmetrical clean armor, instant costume swap, no in-between morph stages, untextured smooth plating, transformation happening off a single cut

Why: transformation shots fail by skipping the middle — the model jumps straight to the finished suit, or renders it factory-clean. You want staged, asymmetric, battle-worn growth. Pair with the Dark Tokusatsu atmosphere prefab.

Combat / impact

weightless floaty motion, no impact, soft contact, characters phasing through each other, missing follow-through, slow-motion drift on every hit, no debris or dust on impact

Why: AI combat reads as weightless — punches that don't land, bodies that don't recoil. Negate the floatiness and demand impact/weight. Pair with the Hollywood Teal-and-Orange or Heavy Mech prefab.

Dialogue / talking head

lip-sync drift, mouth out of sync with audio, frozen dead eyes, mannequin stillness, uncanny face warp on speech, teeth merging, plastic lip texture

Why: close-up dialogue exposes lip-sync drift and the "frozen eyes" uncanny look. Pair with the Commercial Portrait reference workflow so the face is locked before motion.

Environment / wide establishing

warping background geometry, flickering distant detail, melting architecture, repeating tiled textures, impossible perspective, crowd faces smearing, signage gibberish text

Why: wide shots break in the background — smearing crowds, gibberish signage, architecture that boils between frames. Hold the geometry stable.


See also