Guides

How to Record a Clip for a Great AI Avatar

How to record a clip for an AI avatar that looks like the real person: framing, light, sound, what to say and a checklist you can hand to your users.

PitchTwin team, , 6 min read

Most disappointing avatar videos are decided before any API call: the source clip was dark, shaky or noisy. If your product asks users to record themselves, the instructions you show on that screen matter more than any model setting. This guide explains what makes a good source clip and gives you a checklist you can paste into your onboarding.

Why the clip matters so much

PitchTwin builds each avatar from real video of the person and keeps that footage. The skin, hair, lighting, background and the person's own expressions in every generated video come from the clip. That is why results look 1:1 like the person instead of a smoothed, re-generated face, and it is also why a bad clip cannot be rescued later.

Put simply: every generated video inherits the clip. Good light in, good light out. A messy background in, a messy background in every video.

The hard requirements

These come from the API, so validate them in your upload screen before the user hits submit:

  • Length: 10 to 120 seconds.
  • Size: up to 50 MB through the upload endpoint, or up to 200 MB if you pass a URL.
  • Content: one person, facing the camera, talking.

Everything else below is about quality, not validity.

Equipment: a phone is enough

A recent phone camera on a stand is fine. Use the back camera if someone else is filming, or the front camera with the phone on a tripod if the person records alone.

Do not use 4K. PitchTwin normalizes every clip to a long side of at most 1920 px and 25 fps, so 4K adds file size without adding anything to the result, and it makes the 50 MB upload limit easy to hit. Record in 1080p. You also do not need to worry about phone-specific formats: rotation metadata and HEVC are handled on PitchTwin's side.

Framing

  • Pick the orientation of the final video. The avatar keeps the clip's framing. If your users will post vertical videos, record vertical; for a landscape player, record landscape.
  • Head and shoulders. Frame from roughly mid-chest up, with a little space above the head. The face should be large enough to show detail, but the person should not touch the edges of the frame when they move.
  • Eye level, facing the camera. Put the lens at eye height and have the person look into it, not at the screen. A camera below the chin is unflattering in every video that follows.
  • Keep the camera still. A tripod or a stack of books. Handheld shake ends up in every generated video.

Light

Light is the single biggest factor in how real the avatar looks.

  • Face the light source. A window in front of the person, slightly to one side, is the easiest good light there is.
  • Avoid backlight. A bright window behind the person turns the face into a silhouette and the camera's auto-exposure will flicker.
  • Do not mix colors. Daylight from a window plus a warm ceiling bulb gives the face two different tints. Turn one off.
  • Keep it constant. Recording while clouds pass, or with a light that someone switches on halfway through, makes the clip uneven.

Background

Choose something calm and static: a wall, a tidy shelf, an office corner. Avoid moving elements behind the person (people walking past, a screen with video, a fan), because they will move the same way in every generated video. Avoid a background that dates quickly, such as holiday decorations, if the avatar will be used for months.

Remember that one avatar is one look. If your customer wants a formal and a casual version, record two clips and create two avatars.

Sound

The clip's audio may become the person's voice. If the clip has at least 10 seconds of speech, PitchTwin clones the voice from it automatically. So the recording conditions for the video are also the recording conditions for the voice:

  • Record in a quiet room. Close windows, turn off fans and notifications.
  • Soft furnishings help; an empty room with hard walls echoes.
  • A small clip-on microphone near the chest is a cheap upgrade over the phone's built-in mic.

If the clip is noisy or the person barely speaks, record a separate voice sample instead and pass it as voice_url when creating the avatar. One to two minutes of clean speech gives the best clone. If your product always sends its own finished audio, the voice does not matter at all: the avatar will be created with has_voice: false and used in audio mode.

Performance: what to say and how

The person should talk the whole time, naturally, as if explaining something to a colleague. A clip of someone sitting silently or reading stiffly gives the avatar less natural movement to work with.

Good things to say:

  • A short introduction: name, role, what they do.
  • An explanation of something they know well, which keeps them relaxed.
  • A few sentences in the tone of the videos they will generate later (a sales message, a lesson intro).

Things to avoid:

  • Hands in front of the face, hair falling over the mouth, or a microphone that covers the chin.
  • Big gestures and head turns. Moderate, natural movement looks best; leaving the frame or turning sideways does not.
  • Sunglasses, hats with a deep brim, or anything that hides the face.
  • Eating, chewing gum or drinking during the take.

Aim for comfortably more than the 10-second minimum so the person has time to settle in. There is no need to fill the full 120 seconds.

A checklist you can paste into your product

  1. Record in 1080p, not 4K, in the orientation your videos will use.
  2. Put the phone on a stand at eye level, framing head and shoulders.
  3. Face a window or soft light; nothing bright behind you.
  4. Choose a quiet room and a calm, static background.
  5. Look into the lens and talk naturally the whole time.
  6. Keep your face visible: no hands, hair or objects in front of it.
  7. Record 10 to 120 seconds; stay under 50 MB.
  8. Confirm you agree to an avatar being made from this clip.

The last item is not optional. PitchTwin's Terms require every person in an avatar to have agreed to it, so collect and store that agreement as part of the recording flow.

After recording

Upload the clip from your app with POST /uploads and create the avatar with POST /avatars; the full flow, including polling and a test render, is in how to create an AI avatar from a short video. Then watch one test generation at full size before you ship the avatar to the customer. If something looks off, re-record rather than tweak: the clip is the input that matters most.

Try it

Sign up at /signup for 60 free tokens with no card, record a clip on your phone and upload it from the browser in the dashboard. Generate a test video in the Playground, and read the docs or the avatar from video page when you are ready to build the recording flow into your product.

Keep reading