Business

Why AI avatars look plastic, and how to fix it

Why AI avatars look plastic: the causes behind smooth skin, flat light and drifting faces, plus a practical checklist to get results that look like the person.

PitchTwin team, , 6 min read

You show a prospect an AI avatar video and the first comment is not about the script. It is "the face looks a bit off". Skin is too smooth, hair looks painted on, the lighting does not quite match, and the person looks like a polished cousin of themselves. For a B2B product built on avatar video, that reaction costs you trust before the message even lands.

This post explains where the "plastic" look comes from, how to spot it when you evaluate providers, and what you can do about it in your own product.

Why "almost right" feels worse than "clearly fake"

People are very good at reading faces, and small errors on a face that looks nearly real can feel more unsettling than an obvious cartoon. The roboticist Masahiro Mori described this in his 1970 essay on the "uncanny valley": as a figure becomes more humanlike, our affinity for it rises, then drops sharply just before it becomes fully convincing.

For avatar products, the practical takeaway is that a stylized avatar and a convincing one can both work. The middle ground, a face that is almost the person but too smooth, is where viewers get uncomfortable.

Where the plastic look comes from

The face is re-generated instead of kept

Some approaches build an avatar by learning a person's appearance and then generating every frame of the face from scratch. Generation tends to average out fine detail. Pores, small blemishes, stray hairs, and the slight unevenness of real skin are exactly the details that get smoothed away. The result is recognizable, but it reads as retouched.

Starting from a single photo

A photo contains one expression, one head angle, and one lighting setup. Everything else, such as how the person moves, blinks, and shifts their expression, has to be invented. Photo avatars are useful when a photo is all you have, and some providers price them below video-based avatars for that reason. But invented motion is a common source of the "off" feeling.

Lighting that does not belong to the scene

When a face is generated separately from the rest of the frame, its lighting can drift from the background, the neck, and the hair. Viewers rarely name the problem, but they notice it.

Generic expressions

Everyone has their own way of raising their eyebrows, smiling while talking, or tilting their head. Models trained on many people tend to produce an averaged expression set. The avatar ends up looking like the person, but not moving like them.

Drift over time

If a provider updates its model, an avatar can look slightly different from one month to the next. That is not plastic in the strict sense, but it has the same effect: the face stops matching the person your customer's audience knows.

Poor source material

Every method suffers with bad input. Low-resolution clips, heavy compression from messaging apps, harsh overhead light, and beauty filters applied by the phone all reduce detail before the avatar is built. If the source is already smoothed, the result will be too.

How to evaluate a provider for this

Marketing reels are chosen to look good. Test with your own face, or a teammate's, under ordinary conditions.

  1. Record one plain clip on a phone in normal daylight. Do not use a studio setup for the test.
  2. Create an avatar from it with each provider you are considering.
  3. Render the same script of around 30 seconds with each.
  4. Watch at full resolution on a large screen, not a phone preview.
  5. Check specific areas: skin texture on the cheeks and forehead, hair edges against the background, teeth, the lighting on the face compared with the neck, and how blinks look.
  6. Show the results to someone who knows the person and ask which one looks most like them. That is the test your customers' viewers will apply.

Repeat the test a few weeks later with the same avatar to check for drift.

How to fix it

Build from real video and keep the footage

The most direct fix is to avoid re-generating what the camera already captured. PitchTwin builds each avatar from real video of the person and keeps that original footage, so skin, hair, lighting, and the person's own expressions come from the recording itself. The result looks 1:1 like the person, rather than a smooth re-generated face.

Creating an avatar is one request. Upload the clip (or pass a URL), then poll until it is ready:

curl -X POST https://pitchtwin.ai/api/v1/avatars \
  -H "Authorization: Bearer pt_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Sales rep - Alex",
    "video_url": "upload:file_abc123"
  }'

The response is 202 with status processing. Poll GET /avatars/:id until the status is ready. If the clip contains at least 10 seconds of speech, the voice is cloned from it too; for the best voice clone, pass a separate 1 to 2 minute sample as voice_url. The full flow is in how to create an AI avatar from a short video.

Record a better source clip

Because the footage is kept, the clip you record is the clip your viewers see. A few rules make a large difference:

  • Use the phone's rear camera at its highest resolution, and turn off beauty or portrait filters.
  • Face soft, even light, such as a window, rather than an overhead bulb.
  • Send the original file, not a copy that passed through a messaging app.
  • Keep the head roughly still and centered, with natural small movements.
  • Record 10 to 120 seconds; PitchTwin accepts clips in that range (up to 50 MB by upload or 200 MB by URL) and normalizes phone rotation, HEVC, and frame rate automatically.

There is a full guide in how to record a clip that makes a great AI avatar.

Pin the version

A face that looks right today should look right next year. On PitchTwin, each avatar is pinned to its engine version, so there are no silent upgrades and faces do not drift. If you use another provider, ask how they handle model updates for existing avatars.

Do not neglect the voice and pacing

A face that looks real paired with flat, rushed speech still feels synthetic. Natural pacing helps viewers accept the video. Use the script field to add pauses between sentences:

{
  "avatar_id": "av_3f9c1e7a2b4d6e80",
  "script": [
    { "text": "Hi Sam, thanks for taking a look.", "pause_after_ms": 400 },
    { "text": "I recorded this to walk you through the proposal.", "pause_after_ms": 300 },
    { "text": "It should take about a minute.", "pause_after_ms": 0 }
  ],
  "voice": { "speed": 0.95, "language": "en" }
}

More on this in writing scripts for AI avatars.

Summary

CauseFix
Face re-generated, detail averaged awayBuild from real video and keep the footage
Single photo sourceUse a short video clip instead
Mismatched lightingKeep the recorded lighting; record in even light
Generic expressionsKeep the person's own expressions from the clip
Drift after model updatesPin each avatar to an engine version
Low-quality inputRecord a clean, unfiltered, original-resolution clip
Flat deliveryScript pauses and a natural speaking speed

Try it

The fastest way to judge is with your own face. Sign up for 60 free seconds of video with no card, upload a phone clip in the dashboard, and compare the result side by side with the original recording.

Keep reading