How do I create a Tavus Replica from a ~2-minute video, and what should my video look like to get a good result?

Most creators overthink the “perfect” Tavus Replica and under-think the source video. In reality, a clean, simple ~2-minute recording with good lighting and clear audio will get you dramatically better results than a highly produced, over-edited clip. The Replica is learning your face, your expressions, and how you speak—so the more natural and well-captured that is, the more lifelike your AI Human will feel.

Quick Answer: You create a Tavus Replica by uploading a short, well-lit talking-head video (around 2 minutes) where you speak naturally, facing the camera. For the best result, focus on: stable framing, neutral background, consistent lighting, clear audio, and a range of natural expressions while you talk.

The Quick Overview

What It Is: A Tavus Replica is a high-fidelity AI Human modeled on your face and voice, generated from a short reference video and used for real-time, face-to-face AI interactions.
Who It Is For: Builders who want to embed lifelike AI Humans into products, workflows, or demos, and individuals who want a familiar, human-like presence powering their PALs.
Core Problem Solved: Instead of a generic avatar or disembodied voice, you get a trusted, recognizable AI Human that can see, hear, and respond like you—at the speed of real conversation.

How It Works

Under the hood, a Tavus Replica is a training artifact for our rendering stack, not a pre-recorded video template. Your ~2-minute video gives Phoenix-4 (our gaussian-diffusion facial behavior model) enough signal to learn your facial structure, micro-expressions, and the way your mouth actually shapes speech.

From there, when your Replica is live in an AI Human:

Perception: Raven-1 takes in multimodal input—your user’s voice, video, screenshare, and their nonverbal cues.
Understanding & Speech: We run perception → speech recognition → LLM reasoning → TTS, all tuned for sub-second turn-taking.
Real-Time Rendering: Phoenix-4 uses your Replica to render temporally consistent, high-fidelity facial behavior that matches the generated audio and conversational timing guided by Sparrow-1.

You’re not stitching together canned clips. You’re giving the system enough data to synthesize a live, responsive, face-to-face AI Human that feels like talking to someone on a video call.

Step-by-Step: How to Create a Tavus Replica From a ~2-Minute Video

1. Plan a Simple, Talking-Head Setup

You don’t need a studio. You do need clarity.

Camera: Laptop webcam, phone camera, or an external webcam all work.
Framing: Head and shoulders in frame. Centered. No extreme angles.
Orientation: Horizontal (landscape) is ideal for most deployments.
Distance: Roughly arm’s length from the camera—your face should be clearly visible without distortion.

2. Optimize Lighting and Background

Phoenix-4 cares about clear, consistent detail.

Lighting:
- Face a light source (window or soft light) so your features are evenly lit.
- Avoid strong backlight (bright window behind you) and harsh side-shadows.
- Keep lighting consistent for the full ~2 minutes; no flickering or color shifts.
Background:
- Neutral, uncluttered background (plain wall, simple room).
- Avoid moving objects or people behind you.
- No flashing screens, busy patterns, or strong light sources in frame.

3. Capture Clean, Natural Audio

Even though Replica rendering is visual-first, natural speech helps align mouth shapes and expressions.

Environment:
- Quiet room. Close doors and windows.
- Turn off fans, loud AC, and background music.
Microphone:
- Built-in laptop/phone mic is okay if you’re close and the room is quiet.
- A headset or external mic is even better.
Delivery:
- Speak at a normal volume and pace.
- Avoid whispering or shouting.
- Don’t add heavy background music; keep your voice dominant.

4. What to Say in Your ~2-Minute Video

The content matters less than the variety of expressions and phonemes. You’re giving the model a full “sampling” of your face in motion.

Aim for:

Natural speech: Talk like you would on a Zoom call, not like a memorized script.
Varied expressions: Smile, nod, show mild surprise, emphasize a point with your eyebrows—normal, human micro-expressions.
Full mouth movement: Pronounce clearly; don’t mumble.
Steady head position: Small natural movements are good, big head turns are not.

A simple structure you can follow:

Intro (30–40 seconds):
“Hi, I’m [Name]. I’m recording this so Tavus can build an AI version of me. In real life, I work on [role/field], and I mostly talk to people about [topic].”
Story / Explanation (60–80 seconds):
Share a short story, explain what you’re working on, or talk about something you’re passionate about. Let your tone naturally rise and fall.
Closing (20–30 seconds):
“Thanks for watching this recording. I’m excited to see how the AI Human turns out and how it can help people [do X].”

What to avoid:

Reading robotically from a teleprompter.
Overacting or exaggerated expressions that you’d never use in a real conversation.
Extremely fast talking or long silent staring at the camera.

5. Behaviors That Help (and Hurt) the Replica

Helpful:

Gentle nods while speaking.
Natural blinking.
Mild shifts in gaze (but mostly toward the camera).
Subtle hand movement visible in the lower frame (optional).

Problematic:

Turning your head sharply left/right for extended periods.
Covering your mouth (hand, mic, clothing).
Wearing large reflective glasses that hide your eyes.
Constantly looking off-screen or at another monitor.
Chewing gum, eating, or drinking on camera.

Features & Benefits Breakdown

Core Feature	What It Does	Primary Benefit
Short-Form Training (~2 min)	Uses a brief reference video to learn your facial structure and expressions	Fast setup without a studio shoot; get a usable AI Human quickly
Phoenix-4 Facial Rendering	Generates temporally consistent facial behavior in real time	Conversations feel like live video, not stitched-together clips
Multimodal Real-Time Pipeline	Connects perception → ASR → LLM → TTS → AI Human video	Your Replica can see, hear, and respond at the speed of human dialog

Ideal Use Cases

Best for developers building face-to-face agents: Because it lets you embed a white-labeled AI Human into your app with one API, instead of a generic chatbot UI.
Best for individuals creating a familiar PAL: Because your Replica gives your personal AI companion a recognizably “you” presence that feels closer to a real call than a text window.

What Your Video Should Look Like (Concrete Examples)

To match the intent of how-do-i-create-a-tavus-replica-from-a-2-minute-video-and-what-should-my-video-l, here are practical “this vs that” guidelines:

Strong Example:

You sitting at your desk, laptop at eye level.
Window in front of you, soft light on your face.
Plain wall or tidy shelves behind you.
You talking for ~2 minutes about your work, occasionally smiling and nodding.

Weak Example:

You walking outside with your phone.
Sun behind you, your face in shadow, wind noise in the mic.
Cars passing, people crossing the background.
You glancing around, audio cutting in and out.

Strong Example:

You on a video call setup you already use daily.
No filters, no virtual background.
Clear, stable camera, normal body language.

Weak Example:

Heavy beauty filters or AR face effects active.
Aggressive smoothing that removes facial detail.
Virtual background glitching around your hair and shoulders.

Limitations & Considerations

Quality in, quality out: Poor lighting, noisy audio, and unstable framing can limit the realism of your Tavus Replica. If your reference video struggles to see your face clearly, Phoenix-4 has less signal to work with. A quick re-record in better light is usually worth it.
It’s a live AI Human, not prerecorded video: Your Replica doesn’t “play back” your ~2-minute script. It powers real-time rendering tied to live conversation. If you expect pre-scripted, fixed clips, you’re thinking about static avatars, not AI Humans.

Pricing & Plans

Tavus is designed for two main journeys: building with APIs and connecting with personal AI Humans.

Developer Account: Best for developers, founders, and teams who want to embed real-time, face-to-face AI Humans into their own products and workflows. You’ll use Replicas as part of a broader perception → ASR → LLM → TTS → Phoenix-4 pipeline, with enterprise-grade performance, sub-second latency, and white-label options.
PALs Account: Best for individuals who want personal AI companions that listen, remember, and are always present. Your Replica can give that PAL a familiar face and presence, so it feels more like talking to a friend than typing into a chat box.

For current plan details and usage-based pricing, sign up and review the options inside the Tavus platform.

Frequently Asked Questions

How long should my Tavus Replica training video be?

Short Answer: Around 2 minutes of continuous, natural talking-head video is ideal.

Details: You don’t need a 10–20 minute recording. Phoenix-4 is optimized to extract a lot of signal from a compact clip if the conditions are good: stable framing, good lighting, clear audio, and natural expressions. If you’re under 60 seconds, or if half the clip is silence or looking away, consider re-recording to hit ~2 minutes of active speaking time.

Do I need professional gear to get a good Replica?

Short Answer: No. A laptop or phone camera in good light is usually enough.

Details: Presence is an engineering constraint, not a studio budget. The system cares more about signal quality than equipment price. A clean, well-lit, front-facing video from your laptop webcam will generally outperform a 4K DSLR shot with harsh backlight and noise. Prioritize:

Facing a light source.
Quiet environment.
Camera at eye level.
Natural speech for the full ~2 minutes.

If you already have a good video-call setup, that’s usually the perfect place to record your Replica.

Summary

Creating a strong Tavus Replica from a ~2-minute video is less about production value and more about clarity. Sit down, face the camera, light your face, speak naturally, and let your real expressions show up. That’s the raw material Phoenix-4 uses to build a high-fidelity AI Human that can see, hear, and respond in real time.

Get the reference video right once, and you unlock a Replica that can power everything from live AI SDRs to personal PALs—bridging the gap between “chatbot with a face” and genuine, face-to-face conversation.

Next Step

Get Started