Work / Video and content

AI avatar Stas BeryNeParsya: a character, a voice and facial motion with no shoot

What it was

We wanted to check whether an AI avatar can hold attention for 26 seconds: not a still image with a voice on top, but a person who talks, gestures and looks into the camera. The look of a popular blogger was the starting point.

What we did

Generated the character's portrait in Nano Banana, put him in an armchair in a bright living room, voiced him in ElevenLabs and animated speech and gestures. The entry into the frame goes through an empty room and a fiery flash on the hoodie; at the ninth second an Instagram profile screen is inserted.

What the client got

A 26-second monologue in one shot: lips moving with the words, hands gesturing, eyes on the camera. A character who does not exist looks like a blogger in his own living room.

ClientOwn project: an AI version of a popular blogger
My roleCharacter portrait, scene, voice, speech and gesture animation, editing
StatusPublished on YouTube Shorts
Result26 s of monologue in one shot with no live footage · Open
Rejected: a 9-panel collage as the character card, the model does not read the face from it. A reconstruction of a rule from our base, not a screenshot from this project
Rejected: a 9-panel collage as the character card, the model does not read the face from it. A reconstruction of a rule from our base, not a screenshot from this project
Accepted: one clean portrait as the reference, just the head up close. The card is a reconstruction, the frame is real, from the finished video
Accepted: one clean portrait as the reference, just the head up close. The card is a reconstruction, the frame is real, from the finished video

The task

Make an AI version of a popular blogger: not a picture with a voice on top, but a character who sits in an armchair, talks to the camera, gestures and holds attention for 26 seconds. 9:16 for Shorts and Reels. No shoot at all: the face, the room, the voice and the movement are generated.

Why, and what the idea was

An avatar that just moves its lips against a static background gives itself away in the first second. We wanted the character to live in the frame: sit relaxed, lean in, show things with his hands, look you in the eye. So the scene is not a studio but a bright living room with windows, a plant and an armchair, and the entry into the frame goes through an empty room and a flash on the hoodie, so the first second works as a hook.

How it went

When What happened
Stage 1 The character’s portrait in Nano Banana: a young man, dark curly hair, black hoodie, a chain
Stage 2 The scene: an armchair in a bright living room, windows, a plant. Entry into the frame through an empty room and a fiery flash
Stage 3 Voice in ElevenLabs, speech and gesture animation to the finished track
Stage 4 Edit: the Instagram profile screen inserted at the ninth second, delivery at 26 seconds

Tools and method

Portrait: Nano Banana. Voice: ElevenLabs. Animation: Seedance. A detailed log of this job did not survive, so here is only what is visible in the video and the rules we make avatars by now.

The face reference is one clean image: the head up close, no shoulders, no background. No words about race or skin color in the text; the face comes from the photo only. The voice script is written with emotion tags and pauses so the lines do not run in a flat row, while the lip-sync tool gets clean text without tags.

Result

The video is 26 seconds, 9:16, one shot. 0-2 s: an empty living room, a desk with a computer. 2-4 s: the character appears in the armchair, fiery lines flare up on the hoodie and fade. 4-26 s: a monologue to the camera, hands in the frame the whole time; at the ninth second he points at an inserted Instagram profile screen.

Mistakes and what we took from them

A collage as the character card does not work.The model does not read the face from a 9-panel collage and gives a similar but different person. One clean portrait, just the head up close, and front plus profile is more accurate than a single shot.
Race words in the text override the reference.Write a skin color into the prompt and the model draws that, not the person in the photo. Such words are gone from our prompts entirely.
Voice with tags, lips without.In ElevenLabs the text goes in with emotion tags and pauses, otherwise the lines sound flat. The lip-sync tool gets clean text; tags break the result there.

Tools

Nano BananaElevenLabsSeedance