This skill is the flagship of the AI Content cluster. Covers the entire process of creating AI Avatar videos from zero to publish, voice clone, anti-detection, and VN legal standards.
1. For beginners (Newbie Guide)
What is an AI Avatar?
An AI Avatar is a video with your image (or a representative image) but the AI speaks instead. You only need 1 photo or 1 short selfie video — the AI will create a video with voice, gestures, and natural expressions. No filming, no studio, no actor needed.
What do you need to get started?
| Method | Requirement | Quality |
|---|---|---|
| Portrait photo | 1 front-facing photo, clean background, 1024×1024+ | Average — mouth not yet natural |
| Selfie video | 1 30s video, looking straight at camera, speaking naturally | Good — more accurate lipsync |
| Custom avatar | Record 2-5 minutes with teleprompter + external mic | Excellent — almost like a real person |
Minimum equipment: A phone with HD front camera + external mic (or headset with mic).
How long does it take?
- 1 single video (60s): 30-60 minutes (including script + render)
- Batch of 10 videos: 1-2 days
- Batch of 30 videos: 4-5 days (with an optimized process)
Cost?
| Tier | USD/month | VND/month (estimated) | Number of videos |
|---|---|---|---|
| Free | $0 | 0d | 1-3 videos, with watermark |
| Pro | $30-100 | 750K-2.5M | 10-30 videos, no watermark |
| Enterprise | $200-500+ | 5M-12.5M+ | 30+ videos, custom avatar, API |
5 common mistakes when just starting out
- Lipsync out of sync: Cause — script too fast or AI voice not yet matched. Fix: reduce speaking speed by 10-15%, use voice clone instead of the default voice.
- Voice doesn't sound like me: Cause — sample too short or too much background noise. Fix: re-record 3-5 minutes in a quiet room, read a passage containing many different sounds.
- Video flagged as "AI content": Cause — the platform detects AI patterns. Fix: see the Anti-detection section below.
- Video is blurry/soft: Cause — low-quality input image. Fix: use an image of at least 1024×1024, natural lighting, no filter.
- Render takes too long: Cause — free tools have long queues. Fix: render during off-peak hours (early morning VN = night in the US) or upgrade to Pro.
2. Information gathering
Ask at most 4 questions before starting:
- Purpose of use? Brand awareness / Sales / Education / Internal?
- Main platform? TikTok / YouTube / Facebook / Instagram / LinkedIn / Zalo?
- Budget tier? Free ($0) / Pro ($30-100) / Enterprise ($200+)?
- Desired number of videos/month? 1-5 / 10-30 / 30+?
Based on the 4 answers, automatically select the appropriate Tier + Workflow. If the user has already uploaded a reference image, do not ask for a long form; classify the image first, create the prompt/setup, then only ask for the missing assets.
3. Tier decision — Tool & Cost
| Tier | Suggested tool | Price/month | Quality | Limit | Suitable for |
|---|---|---|---|---|---|
| Free | Captions Free, HeyGen Trial, D-ID Trial | $0 | 6/10 — watermark, duration limit | 1-5 videos, max 60s/video | Individual testing, new freelancers |
| Pro | HeyGen Creator ($29), Synthesia Starter ($29), ElevenLabs Pro ($22) | $30-100 | 8/10 — no watermark, HD | 10-30 videos, max 5 minutes/video | SME, small agency, content creator |
| Enterprise | HeyGen Business ($89+), Synthesia Enterprise (custom) | $200-500+ | 9.5/10 — custom avatar, API, priority render | 30+ videos, no limit | Large agency, large brand, e-learning |
Quick suggestion:
- Just starting: HeyGen Trial (1 free video, full experience)
- Serious but low budget: Captions Pro ($10/thang) cho lipsync + ElevenLabs Starter ($5) for voice
- Scale fast: HeyGen Creator + ElevenLabs Pro = best combo for price/quality
- Enterprise VN: Contact Vbee (best Vietnamese voice AI) + HeyGen Business
4. Workflow 1: Single Avatar Production
1 single video — from script to publish in 30-60 minutes.
6-step process
| Step | Task | Suggested tool | Time |
|---|---|---|---|
| 1. Script | Write a 150-300 word script (60s video) | Skill 04-script-video |
10 minutes |
| 2. Voice | Create a voice or use a voice clone | ElevenLabs / Vbee / HeyGen Voice | 5 minutes |
| 3. Avatar | Choose an available avatar or upload an image/video | HeyGen / Synthesia / D-ID | 3 minutes |
| 4. Render | Combine voice + avatar, choose background, gesture | Tool selected in step 3 | 5-15 minutes (render) |
| 5. QA | Check according to the 100-point QA Score (section 12) | Manual review | 5 minutes |
| 6. Publish | Export MP4 → post to platform | Manual / Scheduler | 2 minutes |
Script template for AI Avatar (60s)
[HOOK — 3s] Mot cau gay to mo, dat van de
[PROBLEM — 10s] Mo ta van de khach hang gap
[SOLUTION — 25s] Giai phap cua ban, 2-3 diem chinh
[PROOF — 12s] So lieu, testimonial, ket qua
[CTA — 10s] Hanh dong cu the: "Nhan link bio de..."
Output template
## Video AI Avatar — [Ten san pham] — [Ngay]
- **Script:** [Noi dung]
- **Voice:** [Tool] — [Voice ID / Clone name]
- **Avatar:** [Tool] — [Avatar ID / Custom]
- **Resolution:** 1080×1920 (9:16) hoac 1920×1080 (16:9)
- **Duration:** [X]s
- **QA Score:** [X]/100
- **Publish:** [Nen tang] — [Link]
5. Workflow 2: Multilingual translation
1 original video (VN) → many languages (EN, TH, ID, JA) for the Southeast Asian + international markets.
Use case
- VN brands want to expand to Thailand, Indonesia, Philippines
- Online courses need multilingual captions
- Agencies making videos for international clients
Tool comparison — Translate & Lipsync
| Tool | Languages | Lipsync | Price | Notes |
|---|---|---|---|---|
| Rask AI | 130+ languages | Good (SyncLabs tech) | $50/month (Pro) | Best for translate currently |
| HeyGen Translate | 40+ languages | Decent | Included in Creator+ plan | Built-in integration, convenient |
| Synthesia Translate | 35+ languages | Decent | Included in Enterprise plan | Suitable for e-learning |
Process
- Create the original video in Vietnamese (Workflow 1)
- Upload to a translate tool (Rask AI recommended)
- Choose the target language — the tool automatically translates + lipsyncs
- Review the translation (remember to have a native speaker check it)
- Export and publish for each market
Important notes with Vietnamese:
- Vietnamese lipsync is not yet perfect (complex tones and accent marks)
- Workaround: shoot the original in English → translate back to Vietnamese
- Or: use a Vietnamese voice clone (Vbee) + a separate avatar instead of translating
See details at
references/ai-avatar-tools-vn.mdthe Translate section.
6. Workflow 3: Batch Production
30 videos in 5 days — an assembly line process.
Detailed timeline
| Day | Work | Output | Tool |
|---|---|---|---|
| Day 1 | Script batch — write 10 scripts based on a template | 10 scripts (.md) | Skill 04-script-video + AI assist |
| Day 2 | Voice batch — render 10 audio files | 10 audio (.mp3) | ElevenLabs API / Vbee API |
| Day 3 | Avatar batch — upload audio + avatar, queue render | 10 videos rendering | HeyGen Batch / Synthesia |
| Day 4 | QA batch — review 10 videos, fix errors, re-render | 10 QA'd videos | Manual + QA Score |
| Day 5 | Publish batch — export, add captions, schedule posts | 10 videos published | Buffer / Later / Manual |
Repeat for 3 consecutive weeks = 30 videos. Or scale Day 1-2 up to 15 scripts/week.
Estimated cost for a batch of 30 videos/month
| Tier | Tool combo | Cost/month | Cost/video |
|---|---|---|---|
| Free | HeyGen Trial + Captions Free | $0 (limited to 3-5 videos) | $0 (with watermark) |
| Pro | HeyGen Creator + ElevenLabs Pro | ~$51 (~1.3M VND) | ~$1.7 (~42K VND) |
| Enterprise | HeyGen Business + ElevenLabs Scale | ~$189 (~4.7M VND) | ~$6.3 (~157K VND) |
Tips for batch optimization
- Script template available: Create 3-5 script frameworks, only change the core content
- Consistent voice: Use 1 voice clone for the entire series
- Render during off-peak hours: Queue render from 22h-6h (VN) to reduce waiting time
- QA checklist: Print the QA Score on paper, check each video like an assembly line
7. Workflow 4: Hybrid Real + AI
Combine a real face (actual filming) + AI body — increase trust and reduce production time.
Use case
- Format 1: Real face intro 5s + AI body 55s (save filming time)
- Format 2: AI video on weekdays + Real video on weekends (balance quality/effort)
- Format 3: Real talking head + AI B-roll (professional like a studio)
Hybrid editing technique
- Shoot a real intro: 5-10s, look straight at the camera, greet naturally
- Create the AI for the rest: Use the same outfit, same background (or green screen)
- Edit in CapCut/Premiere: Cut precisely, add smooth transitions
- Color match: Adjust the AI color to match real footage (LUT or manual)
Suitable tools
| Part | Tool | Note |
|---|---|---|
| Real face lipsync | Captions | Fix speech errors in real video |
| AI body | HeyGen / Synthesia | Match outfit with real footage |
| Editing | CapCut (free) / Premiere | Smooth transitions, color match |
| Color grade | CapCut filter / DaVinci Resolve (free) | Match tone between real and AI |
Advice: Hybrid significantly increases trust — customers see a real face at the start of the video then AI for the rest is very smooth. Engagement effectiveness increases 20-35% compared to full AI.
8. Voice Clone Protocol
Voice sample requirements
| Criteria | Requirement | Why |
|---|---|---|
| Duration | 3-5 minutes | Enough data for AI to learn the voice |
| Quality | WAV/FLAC, 44.1kHz+, mono | Reduce noise, increase accuracy |
| Environment | Quiet room, no echo | Noise reduces clone quality |
| Read content | Text with many different sounds (a, e, o, u, i) | AI needs to hear all sound ranges |
| Emotion | Read naturally, casually, without being stiff | The best "neutral" voice AI clone |
Tool comparison — Voice Clone
| Tool | Best for | Voice clone price | VN quality | Notes |
|---|---|---|---|---|
| ElevenLabs | English, multilingual | From $5/month (Starter) | 7/10 | Best overall, VN decent |
| Vbee | Vietnamese | Contact for pricing | 9/10 | Best for Vietnamese |
| HeyGen Voice | Used within HeyGen | Included in Creator+ plan | 6/10 | Convenient if you already use HeyGen |
| Resemble AI | Enterprise, API | From $99/month | 6/10 | Strong API, decent quality |
Consent form template
MANDATORY before cloning anyone's voice.
DONG Y SU DUNG GIONG NOI
Toi, [HO TEN], dong y cho [TEN CONG TY/APP] su dung giong noi
cua toi cho muc dich: [MO TA CU THE — vd: tao video marketing
cho san pham X tren nen tang Y].
Thoi han: [X thang / Vo thoi han / Den khi rut dong y]
Ngay ky: [DD/MM/YYYY]
Chu ky: _______________
Reference: View details
references/voice-clone-prompts-vn.md
9. Avatar Setup Checklist
Before recording/uploading video or images for an AI avatar:
- Lighting: Natural light or softbox, no harsh shadows on the face
- Background: Solid color background (white/gray) or real background (office, store)
- Clothing: Suitable for the brand, avoid small striped patterns (causes AI noise)
- Framing: From the chest up, eyes at the top 1/3 of the frame
- Eye contact: Look straight into the camera (don't look at the screen)
- Gestures: Natural, hands can rest on the table or gesture lightly
- Resolution: At least 1080p (1920×1080), 4K recommended
- Aspect ratio: 9:16 (TikTok/Reels), 16:9 (YouTube), 1:1 (Feed)
- File format: MP4 (H.264) for video, PNG/JPG for images
- Backup: Save original files to the cloud (Google Drive / OneDrive) before uploading to the tool
9A. Reference Image -> Avatar Prompt Director
Use when the user throws in one or more reference images and wants to create an avatar, swap faces, change brand colors, add a logo, or create a prompt before uploading to the tool.
Classify input images
| Image type | Role | Requirements |
|---|---|---|
| Style ref | Capture mood, lighting, background, outfit, camera angle | Do not use for the face unless the user explicitly says so |
| Face ref | Keep identity / swap face | 1-3 clear face photos, no filter, front + 3/4 |
| Video selfie | Create a more natural custom avatar/lipsync | 30s-2 minutes, look at the camera, speak naturally |
| Logo/palette | Attach personal/company brand | PNG/SVG logo + 2-4 hex colors |
| Product/location | Make props/backdrop for the avatar | Choose the right image or the right context to use |
Multiple images = multiple streams
## Luong avatar
| Flow | Anh dau vao | Vai tro | Tool goi y | Asset con thieu |
|------|-------------|---------|------------|-----------------|
| A | style-01 | style/background | 30-thiet-ke-master -> HeyGen | face ref, logo |
| B | face-01 | identity | HeyGen custom avatar | script, voice sample |
- If each image is a different style, create a separate prompt for each Flow.
- If multiple images serve the same avatar, group them by role: style + face + logo + palette + product.
- Clearly request each next asset: face image, selfie video, logo, hex code, script, voice sample.
Prompt setup output
## Avatar prompt setup — Flow A
- Style ref:
- Face ref:
- Brand assets:
- Target platform:
- Tool route:
## Copy-paste visual prompt
[English prompt for avatar/source image generation]
## Upload next
- Face/video selfie:
- Logo:
- Brand colors:
- Voice sample:
- Script:
If you only need a static avatar image, route to 30-thiet-ke-master personal-brand mode. If you need a talking-head video, continue this workflow.
10. Anti-detection for FB/IG/TikTok VN
5 signs AI gets detected and how to fix them
| Sign | Platform often catches it | How to fix |
|---|---|---|
| Face is "stiff", no natural blinking | FB, IG | Use a selfie video instead of a photo; choose an avatar with micro-expressions |
| Voice is flat, no pauses | TikTok, FB | Use voice clone (with natural pauses) instead of default TTS |
| Background is completely static | FB, IG | Add slight grain, or use a real background |
| Isolated movement (only the mouth) | TikTok | Choose an avatar with gestures (hands, head); use HeyGen v3+ |
| Metadata contains AI tool tags | YouTube (monetize) | Export with CapCut (removes original metadata); add color grade |
Techniques to increase "human feel"
- Add film grain/noise: 2-5% noise in CapCut or Premiere
- Zoom and crop: Zoom in 5-10% and add slight movement (Ken Burns)
- Color grade: Apply a film LUT or adjust manually — avoid colors that are "too clean"
- Text overlay: Add subtitles, callouts, stickers to cover weak AI areas
- B-roll insert: Insert 2-3 b-roll clips (product, lifestyle) every 15-20s
- Sound design: Add background music + light SFX (increase immersion, mask AI voice)
By platform
- TikTok: Quite lenient — prioritize good content over AI checks
- Facebook/Instagram: Fairly strict — needs careful anti-detection
- LinkedIn: Almost no detection — best suited for AI avatars
- YouTube: Strict for monetized videos — must disclose per YPP policy
WARNING: NEVER use an AI avatar to impersonate someone else. This violates the law and will result in permanent account suspension.
11. Ethics & Disclosure VN
Legal framework 2024-2025
| Document | Main requirement |
|---|---|
| Decree 147/2024/ND-CP | Mandatory disclosure when using AI for commercial advertising |
| Advertising Law 2012 (amended 2024) | Prohibits misleading information about the origin of products/services |
| YouTube YPP Policy | Mandatory labeling "Altered content" if using realistic AI |
3-tier disclosure
| Tier | When | Example |
|---|---|---|
| MANDATORY | Commercial advertising (paid ads, paid KOLs) | "This video uses AI Avatar technology" |
| SHOULD HAVE | Regular content (organic post, educational video) | "Images/voice supported by AI" |
| OPTIONAL | Internal test, prototype, demo | Not mandatory but should be noted |
Disclosure template (copy-paste)
Tieng Viet: "Video nay su dung cong nghe AI Avatar de tao hinh anh va giong noi."
Tieng Anh: "This video uses AI Avatar technology for visuals and voice."
Placement: Video description, first 3 seconds of video (text overlay), or the platform's "AI Generated" tag (if available).
Reference: See full
references/ai-video-disclosure-vn.md
12. QA Score — 100 points
Scoring table
| # | Criteria | Points | Description |
|---|---|---|---|
| 1 | Lipsync | /10 | Mouth matches speech, no offset over 0.2s |
| 2 | Voice match | /10 | Voice sounds like a real person (if cloned) or natural (if TTS) |
| 3 | Visual quality | /10 | Image is sharp, no artifacts, no blur |
| 4 | Background | /10 | Background fits the context, no render errors |
| 5 | Lighting | /10 | Even lighting, no odd shadows, matches the background |
| 6 | Gesture | /10 | Natural gestures, not jerky, with hand/head movement |
| 7 | Script flow | /10 | Coherent content: Hook → Problem → Solution → CTA |
| 8 | Disclosure | /10 | AI label attached in compliance (section 11) |
| 9 | Platform fit | /10 | Correct aspect ratio, duration, format of the platform |
| 10 | CTA | /10 | Clear call to action, easy to perform |
Scoring guide
| Level | Points | Action |
|---|---|---|
| Excellent | 90-100 | Publish immediately |
| Good | 70-89 | Can publish, note improvements for next time |
| Needs fix | 50-69 | Fix items below 7 points then re-render |
| Redo | <50 | Review script + voice + avatar from scratch |
Reference
references/ai-avatar-tools-vn.md— Detailed comparison of tools by each criterionreferences/voice-clone-prompts-vn.md— Prompts and sample scripts for voice cloningreferences/ai-video-disclosure-vn.md— Full legal guidance
Skill chain: This skill is often called together with
04-script-video(write script) and25-voice-clone-podcast(clone voice). See CLAUDE.md for the full workflow.Skill 24 | Over Powers Agency | v1.1.0