AI 影片提示詞為何會令手部變形、面孔閃爍,以及如何修正
Why your AI video prompts produce morphing hands and flickering faces (and the syntax that fixes it)
Motaz Hefny 指出,AI 影片提示詞若未明確分配鏡頭、主體動作、環境、光學效果及時長,容易出現鏡頭漂移、手部變形和麪孔跨影格變化。他建議每個鏡頭只保留一項主要動作,並明確加入面部一致、手部形態穩定及影格平滑等限制,取代泛泛的質素形容詞。不同引擎亦需調整寫法,例如 Kling v1.5/v2.0 應在主提示詞正面陳述限制,Runway Gen-3/4 則每個鏡頭最多使用一個鏡頭動詞。
Most AI video generators fail for the same three reasons: the camera moves when it should not, hands melt, and faces shift between frames. I have spent the last few months building prompt frameworks for Kling, Runway, Sora, and Luma, and the failures are not random. They follow patterns.
1. Video prompts are shot briefs, not sentences
Text models read a prompt as one stream. Video models decompose it into shots: camera, subject, motion, lighting, duration. If your prompt reads as a sentence, the model guesses the shot structure. If it reads as a brief, the model executes it.
Weak:
A man walking in a rainy street at night, cinematic, beautiful.
The camera is undefined, so it drifts. "Beautiful" is undefined, so it samples. Better:
Close-up, locked camera. A man in a dark coat walks toward the lens on a wet asphalt street, neon reflections. Shallow depth of field, 35mm look. Slow push-in only. Rain in focus. 5 seconds.
Every verb is assigned to exactly one layer: camera, subject, environment, optics, duration.
2. Motion must be budgeted — one primary movement per shot
The number one cause of morphing is competing motion. If the subject walks AND the camera pans AND the background crowds move, the model spends its motion budget on all three and none stays stable.
The rule I apply in the CINE framework:
- One primary motion (subject or camera, never both).
- Secondary motion limited to environment (rain, steam, hair).
- Everything else frozen.
3. Anti-morphing constraints are syntax, not adjectives
"High quality" does nothing. Specific negative constraints do. What actually holds faces and hands together:
- Facial integrity: "consistent facial features across all frames, no identity drift"
- Anatomy guard: "stable hand anatomy, five fingers, no merging of limbs"
- Temporal smoothing: "smooth temporal interpolation, no flicker or frame popping"
- Anti-flicker: "consistent lighting across frames, no exposure jumps"
On Kling and Runway these go in the negative-prompt field or as explicit constraints in the prompt. On Sora-style single-field models, they must be stated positively: "the face remains identical in every frame."
4. Engine-specific tuning
Each engine has quirks worth knowing:
| Engine | Quirk | Fix |
|---|---|---|
| Kling (v1.5/v2.0) | Ignores negative prompts partially | State constraints positively in the main prompt |
| Runway Gen-3/4 | Drifts camera if two verbs | One camera verb max |
| Luma | Temporal flicker in fast motion | "smooth slow motion" + duration cap |
| Sora/Veo | Over-interprets "cinematic" | Specify lens and lighting explicitly |
5. The three frameworks
I built this into a small tool — Fahhim (فهّم) — as three separate frameworks because the failure modes differ:
- CINE — cinematic scenes: camera, lens, lighting, motion budget.
- I2V — image-to-video animation: locks the input image's identity, animates only what you specify.
- SHOT — commercials and social video: hook in frame 1, fast cuts, aspect ratio, on-screen text zones.
The app is free, works offline, and keeps your prompts on the device — which matters if your prompts contain unreleased product footage descriptions or client names.
- App: https://play.google.com/store/apps/details?id=com.prompti.app
- Web version and prompt library: https://fahhim.moteklab.com
If you are building video prompts today, the single highest-leverage change: write the camera instruction first, then forbid secondary motion. That one move fixes more flicker than any "cinematic" adjective.
來源:DEV Community · dev.to