原文要点

After a lot of experimenting, failing, fixing, and testing again, these are the exact steps we ended up following to get **stable**, **cinematic**, and **character-consistent** VEO outputs. If you're trying to make AI-generated videos with VEO (or Sora, Runway, Pika etc.), you’ve probably hit the two biggest pain points: 1. Prompts that are too vague → random, off-brand shots 2. Characters changing faces, outfits, or entire species every scene If you’re starting out, this is the version we wish we had on Day 1. # 1. Start with a real script (don’t prompt VEO first) The *script* is your blueprint — every prompt, character detail, camera move, and scene length comes from here. Your script should follow a clean arc: * Hook + Problem * Brand Solution * Transformation * CTA Keep it visually descriptive like a film director would, rather than a blog writer. # 2. Break your script into 6–8 second scenes Veo caps clips at \~8 seconds. So you’re not making one video. For a 60 second v

原文完整内容

After a lot of experimenting, failing, fixing, and testing again, these are the exact steps we ended up following to get stable, cinematic, and character-consistent VEO outputs.

If you're trying to make AI-generated videos with VEO (or Sora, Runway, Pika etc.), you’ve probably hit the two biggest pain points:

1. Prompts that are too vague → random, off-brand shots 2. Characters changing faces, outfits, or entire species every scene

If you’re starting out, this is the version we wish we had on Day 1.

1. Start with a real script (don’t prompt VEO first)

The script is your blueprint — every prompt, character detail, camera move, and scene length comes from here.

Your script should follow a clean arc:

  • Hook + Problem
  • Brand Solution
  • Transformation
  • CTA

Keep it visually descriptive like a film director would, rather than a blog writer.

2. Break your script into 6–8 second scenes

Veo caps clips at \~8 seconds. So you’re not making one video. For a 60 second video, you’re making 8–12 tiny videos that will later be stitched together.

Each scene must have:

  • One clear visual message
  • A matching chunk of voiceover
  • A clear sense of pacing (you’ll sync it with VO later)

Sample formatting:

SCENE 1: A stressed marketer with 10 tabs open, chaotic lighting. VO: "Managing campaigns feels like juggling fire." Duration: 7 sec

3. Convert each scene into a JSON prompt for VEO

This is the game changer. Plain text prompts → too ambiguous. JSON prompts → precise, structured, and consistent.

JSON format:

{ "prompt": "Detailed visual description with setting, lighting, mood, environment", "duration": 8, "style": "cinematic, brand commercial, soft gradients, natural skin tones", "camera": "slow dolly forward", "character_description": "24-year-old slim male, brown skin, short wavy hair, white t-shirt, expressive eyes, warm confident demeanor" }

Every field matters:

  • prompt → the world you're creating
  • style → overall aesthetic
  • camera → motion (static / dolly / pan / zoom)
  • character\_description → the key to consistency

This structure reduces randomness by \~70%.

4. Maintaining character consistency (the hardest part)

Veo likes to give you a new human every time. Here’s how you stop that:

Method 1: “Add to Scene” (use this first)

Extends the previous clip’s character into the next clip.

Method 2: Upload a reference image

If you have a spokesperson or mascot → this is gold.

Method 3: Repeat the exact character descriptor in every prompt

This is crucial.

Use a high-specificity string like:

>

Copy-paste that into every JSON block. Repetition trains the model.

5. Generate 2–3 variations of every clip

Never trust the first output. Choose based on:

  • Face match
  • Clothing match
  • Skin tone match
  • Lighting continuity
  • Smoothness of motion
  • Brand vibe

If something is off → tighten the JSON (especially lighting + camera).

6. Finish everything in CapCut/Premiere

Stitch the clips in sequence → sync VO (11labs is a good tool) → add subtle zooms → add music → export 1080p.

We’re still experimenting, so if you’ve found tricks, hacks, or better ways to keep characters consistent, please drop them below.

Would love to learn from what the rest of you are discovering.

原帖关键讨论

National_Run_3333 1 points

هذا هو البرومبت السينمائي النهائي، جاهز للنسخ في Google Veo 3 أو أي مولد فيديو مشابه. قمت بصياغته بأسلوب "مشاهد متتابعة" لأن هذا النوع من البرومبتات يعطي نتائج أكثر ثباتًا وتحكمًا في الكاميرا والحركة. � Google DeepMind +١ TAFASIEL forDetails – 30 Second Cinematic Commercial Duration: 30 seconds Style: Ultra-realistic, luxury commercial, cinematic, 4K HDR, 60 FPS, anamorphic lens, premium lighting, realistic reflections, dramatic color grading, smooth camera movement, professional architecture visualization. Scene 1 (0–4s) Wide aerial drone shot over a modern city during golden hour. The camera slowly flies toward a luxurious commercial building with a premium off-white ACP (Aluminum Composite Panel) cladding facade. Hidden warm LED lighting glows around the facade. Large glass entrance. Orange and white branding. People walking naturally. Text on screen واجهة مشروعك... هي أول انطباع يدوم Scene 2 (4–8s) Close-up cinematic shots. Beautiful off-white cladding panels. Perfect aluminum joints. Premium glass reflections. Luxury 3D orange logo TAFASIEL forDetails Macro details. Slow camera slider movement. Soft cinematic lighting. Voice Over: "كل تفصيلة... بتعكس هوية مشروعك." Scene 3 (8–14s) Professional engineers measuring the building. Architect working on a large 3D model. CNC machine cutting ACP panels. Workers installing cladding. Installing illuminated channel letters. LED lighting test. Fast cinematic transitions. Text: تصميم تصنيع تنفيذ Scene 4 (14–21s) The building transforms into its final luxurious appearance. Modern off-white cladding. Orange illuminated logo. Large glass facade. Camera performs a smooth 180-degree orbit. Blue-hour lighting transitions into elegant night lighting. Premium architectural photography. Scene 5 (21–27s) Families and business owners arrive. Customers admire the facade. People take photos. Close-ups of: Luxury details LED lighting Premium finishing Glass reflections Voice Over: "جودة عالية... تنفيذ احترافي... وتشطيب يليق باسمك." Final Scene (27–30s) Front hero shot. Camera slowly pushes in. Logo illuminated. Elegant glow. Company name: TAFASIEL forDetails Subtitle: مع رمضان عبد الرحيم Phone number: 01001506089 Voice Over: "مع رمضان عبد الرحيم... كل تفصيلة تصنع الفرق." Fade to black. Background Music Epic cinematic corporate music. Luxury commercial soundtrack. Soft piano. Modern orchestra. Powerful ending. Natural ambient city sounds. Negative Prompt No text errors. No distorted hands. No low-quality textures. No unrealistic reflections. No extra fingers. No duplicate people. No blurry faces. No watermark. No logo distortion. Perfect architecture. Consistent building design. Consistent company branding. إذا كنت ستستخدم Google Veo 3، فمن الأفضل تقسيم الإعلان إلى 5 مقاطع مدة كل منها 6–8 ثوانٍ ثم دمجها في المونتاج، لأن ذلك يساعد في الحفاظ على ثبات المشاهد وجودة الإخراج. � reddit.com +١

SideMiserable573 1 points

You're very welcome. This process was for a brand launch video, which if I shared would then make you think it's an ad. But since you seem to be curious if any actual video was produced, you can watch it here! [https://youtu.be/4u49l4K3TUI?si=3WJxuEi1bFURCUiS](https://youtu.be/4u49l4K3TUI?si=3WJxuEi1bFURCUiS)

RecognitionFrosty757 1 points

Excellent article. Is it possible to add concrete examples for the Character consistency?

Nervous-North2806 1 points

Nice! What about voice?

deseipel 1 points

Where do you upload character reference images in Flow?

来源：https://www.reddit.com/r/VEO3/comments/1p0slkg/a_guide_on_how_to_craft_veo_prompts_keep/