Preface

The previous article shared tips for writing video generation prompts. After it was published, discussion groups were buzzing. One reader half-joked: "Since videos generated with prompts look this good, can I get in and act too?"

Turns out, the answer is: Yes! Using Doubao.

The first time the author generated a video with this feature, they stared at the screen for a long time: the person in the frame was me, the hairstyle was mine, even the habit of slightly pressing my lips together when I speak was there. But in reality that day, I was still in pajamas, hair messy, never sat in front of the camera.

How is this done? Follow along to see.

1. A Digital Twin Is Not Face Swap

A digital twin creates a dedicated AI avatar that can appear on your behalf, based on your facial and voice features. With Doubao, you record a short video: turn left, turn right, read a few numbers, and smile naturally. The system captures your facial keypoints, bone proportions, micro-expressions, muscle dynamics, and voice timbre and rhythm. Then the Seedance model behind Doubao uses real-time biometric capture, 3D facial modeling, and voiceprint feature transfer to build a dedicated avatar model.

When generating a video, if you choose to use the avatar, AI drives this avatar model to say lines, make expressions, and move. So what you hand over is not just a photo, but a "personal image usage manual."

Someone asks: "Is a digital twin the same as AI face swap?" Of course not. AI face swap means the video exists first, then the face is replaced: swap your face into each frame. The result is that the body, lip sync, and emotions belong to someone else. The face is like a mask; the more movement, the more likely it falls off. Side profiles, occlusion, laughing, and quick head turns are all dead giveaways. A digital twin, by contrast, is you first, then generate video: AI builds your virtual image feature library, then generates video based on that library and your prompt. The two are completely different.

Now Douyin has all sorts of digital twin videos, like the popular "cute baby fights monster" that many have seen. As long as you build a virtual avatar for the child around you, you can replicate the scenes in the video, and the effect is lifelike. Can you go beyond pre-made templates and customize your own scenes? The answer is yes, keep reading.

2. Tips for Building a "Good-Looking Avatar" Before Generating One

Before making a digital twin, the author wants to highlight some techniques. Many avatars look fake; the problem often lies not in the model but in the source material. Imagine: AI needs to build your feature library, and your short recording has top light turning your eye sockets into black holes, cluttered background, thick makeup like a mask, and you're so nervous you barely open your mouth. Can that produce a good video?

2.1 Technique 1: Make It Look Good

Prefer soft natural light. Between 9–11 AM and 3–5 PM, the light is most stable: low angle, moderate color temperature, with natural light–shadow transitions on the face. Indoors, face a window so soft light evenly covers your face.

Makeup should suit the role, not be as refined as possible. For ancient costume videos, emphasize a translucent base, thin eyebrows, and lip tinting. For workplace videos, emphasize a defined brow arch; darker lipstick reads more capable. Men can darken their eyebrows and keep everything clean. Regardless of gender, don't pile on makeup; only by preserving skin texture can AI recognize your expressions.

Hairstyle also matters: for ancient costumes, pull hair fully up to expose the forehead and ears; for modern characters, bangs are fine but shouldn't cover your eyes.

Smile naturally, keep your mouth slightly closed, and don't show your teeth in a big laugh—the mouth shape gets distorted. Look directly at the camera, don't let your gaze wander.

Remember: you are giving AI a complete face, not a heavily retouched face. That's how to get a realistic video.

2.2 Technique 2: Make It Sound Good

For realism, voice is just as important. Record in a quiet, enclosed room, away from traffic and wind noise. Don't shout; slowly tighten your abdomen and let your breath carry each word evenly. Men with strong chest resonance should lower their pitch slightly; women can use a medium-high pitch, but not piercing. More importantly, focus on intonation: pause naturally according to meaning, stress key words, and find places to breathe in long sentences. Human ears take these details for granted, but they are how AI learns "how you speak." The more rhythmic your sample, the less the final speech synthesis sounds like a machine reading a script.

2.3 Technique 3: Keep the Environment Consistent

This rule mainly applies after creating the avatar: you can add selfies to make the avatar more like you. But keep consistency in the photos—don't have one natural and another with retouching filters. Keep clothing, hairstyle, and scene tones fixed, so the generated video character stays consistent and more realistic.

3. Step-by-Step: Create an Avatar with Doubao and Get It on Camera

Doubao's digital twin is currently only available on mobile and has some restrictions: only you can create one, and you must pass liveness verification. You cannot upload someone else's face, unauthorized IP images, or realistic photos or videos of real people. Before starting, update Doubao to the latest version.

Step 1: Enter the create-avatar page

1. Open the Doubao mobile app, go to "New Conversation," and tap "Skills."
2. Tap "AI Creation," then "Video Generation."
3. Tap "Create Avatar."

Step 2: Record the video and finish creating the avatar

1. On the "Record and create AI avatar" screen, choose "Beauty" or "Natural," then tap "Start Creating."
2. Read the specified numbers on the screen, such as "579, 410."
3. After recording, tap "Create Now."
4. Return to the conversation screen and you'll see "Avatar creating." Wait a moment, and it's done.

Step 3: Add selfies to make the avatar more realistic (optional)

1. After the avatar is created, on the conversation page tap "Use Avatar on Camera."
2. Tap the avatar image.
3. Tap "Avatar Materials."
4. Tap "Add Selfies."
5. On the add-selfies page, upload your own photos. Best to include front, left 3/4, right 3/4, and side views, keeping the environment consistent.

Step 4: Get the avatar on camera

In a new conversation, tap "Skills," then "AI Creation," then "Video Generation." Turn on "Use Avatar on Camera," and then you can write your prompt. Here is the author's example prompt:

4.4:
0-2s close-up|CEO sits at desk, head down, flipping through documents, pinching brow, slightly tired
Sound: quiet office, rustling paper
2-5s medium|Female secretary walks over with a stack of files, places them neatly on another desk in the office, movements crisp
5-8s close-up + profile|CEO looks up at secretary
Secretary: Mr. Gu, your files are here.
8-10s close-up|CEO nods slightly
CEO: OK.

The generated effect is as follows. For privacy reasons, the author won't show their own avatar, but in real testing, the AI avatar is extremely realistic, almost indistinguishable from a real person. You should definitely try it.

Step 5: Rebuild the avatar

If you're not satisfied with the avatar you built, it's likely the source material wasn't good enough. Go to avatar management, delete the avatar, and create it again.

Note: Before publishing generated avatar videos, set the avatar authorization scope to "Only Me." Check the lines for sensitive words, homophones, and controversial topics. Add an AI-generated content label. Platforms are strictly enforcing now; don't risk your account to skip a step.

That's all for this article. What are you waiting for? Use your avatar to star in the scenes you've always dreamed of.

4. Summary

A digital twin is not face swap; it builds your exclusive image first, then generates video. To make a "good-looking avatar," the key lies in source material: soft light, moderate makeup, clean hair, steady rhythmic voice, and consistent environment when adding selfies. After creating it on the Doubao mobile app, turn on "Use Avatar on Camera," write a storyboard prompt, and let your avatar act for you. If unsatisfied, you can rebuild. Before publishing, set your authorization, check sensitive words, and add an AI label. What are you waiting for? Go create your own avatar and star in the scene you've always wanted—it's really fun!

Welcome readers who love large models and agent technology to follow the author's Juejin account and column. Future AI video updates and shares will be synced here too. The author's main focus is on frontier research in large models and in-depth exploration. You can also follow the author's WeChat public account 大模型真好玩 to have fun with large models together.