Can You Face Swap Into an Anime or Cartoon Character Live?

People ask this two ways. Some want to show up on a call as an anime character. Others want the reverse: "anime to human", what a favourite character would look like as a real person. A live face swap does the second one, whatever face you give it.

Will a live face swap accept a cartoon face?

Yes. All six of our faces were accepted, including the flat vector cartoon. Before a swap starts, the service looks for a face in the photo you picked and turns it into an identity code; if it can't find one, it says no face was found. It found a face in every one, and every one of the 96 camera frames came out swapped.

How sure it was varied a lot. The face detector scores each find from 0 to 1. The photo, the painting and the soft anime face scored about 0.9; the cel-shaded anime face 0.76; the flat cartoon only 0.53. A low score still counts as a face, but it's a sign the swap has less to work with.

What do you look like after the swap?

Like a real person, every time. A face swap model redraws your face with the identity of the face you picked, using your camera's skin, lighting, expression and head angle. It doesn't copy line art, flat colour or the oversized proportions of a cartoon. What carries over are the features: the 3D cartoon's big round eyes and worried brows, the flat cartoon's full beard.

The "looks like" column is ArcFace similarity between the swapped face and the face we picked, averaged over the clip; higher means more of that face came through. The 3D cartoon scored highest of all, even above the photo, because its features are so pronounced. Both anime faces scored far lower: anime drawings reduce a nose to a dot and a mouth to a line, so there is little for the model to copy.

Why do anime faces work worst?

Because a face swap learns from real faces. The identity code describes the shape of the eyes, nose, mouth and jaw as they'd appear on a real person, and anime style leaves most of that out or exaggerates it: huge eyes, almost no nose, a pointed chin. The model filled the gaps with guesses, and in our frames the guesses showed as dark patches on the cheeks and a visible edge along the jaw. We left those frames out of the pictures here because they don't look good on a call either.

Which cartoon styles make the best face for a live swap?

• 3D cartoon (animated film style): the strongest result in our test. It reads as a real face with exaggerated eyes and brows, which is often exactly the fun people want.
• Semi-realistic painting: as clean as a photo, and an easy way to make a face nobody has seen before.
• Flat cartoon with clear features: works, but the detector was least sure of it. A beard, glasses or a strong brow line give it something to hold on to.
• Anime: accepted, but the weakest likeness and visible patches. Use a realistic version of the character instead.

Whatever the style, the rules for a good source face are the same as for a photo: one face, looking straight at you, eyes open, nothing over it, even light. Character Generator makes faces that way on purpose. And pick a face that suits your own hair, because the swap changes your face but not your hairline.

How do you appear as a drawn character on a call?

Use an avatar instead of a face swap. An avatar is a drawn or 3D character that copies your head movement and mouth, so it stays a drawing the whole time. VTuber Maker moves a 3D anime-style model with your face from a webcam, and PNGTuber Maker animates a PNG avatar from your microphone. Both are free and run in the browser, and with the Facecam extension either one can be your camera in Google Meet, Zoom on the web or Discord in Chrome.

Live face swap is the other half: you become anyone you want, any face you upload or generate, as a real-looking person who moves exactly like you. It's on every paid plan, from $10 a month.

What didn't we test?

We ran the live service's code on a cloud graphics card with a recorded clip, not a live call through the extension, and without the faster TensorRT engines production uses (their output matches within rounding). One camera clip of one man, smiling and facing the camera; a turned head or a woman's face may change the ranking. Six faces is a small sample, all made by one image model. ArcFace similarity measures likeness the way a face recognition model sees it, and it tends to favour the swap model we use, which was trained against it.