Real avatar videos, made with GoodAvatar
Every clip below was generated here
These are not stock renders or agency reels. Each video was produced by the same pipeline you get: one AI-generated portrait, one voice track, one generation run, no editing pass afterwards. They are vertical 9:16 because that is where most short video is watched, and they range from a nine second hook to a seventy second narration so you can judge how the model holds up at different lengths.
Short social hook (9s)
A nine second opener built for the first three seconds of a feed. One portrait, one line of voiceover, no cuts. This is the cheapest thing you can make here: nine credits, and it finishes in a few minutes. Use it to test whether a face and a sentence are enough to stop a scroll before you commit to a longer script.
Product explainer (51s)
A fifty-one second feature walkthrough read straight to camera. The portrait stays locked while the voice carries the explanation, which is what makes this format work for landing pages: no jump cuts, no re-recording when the copy changes. Written as a script, spoken by TTS, generated once. Fifty-one credits total.
Course introduction (51s)
The same fifty-one second shape, used the way a teacher would: a welcome, what the lesson covers, and what to do next. Because the face is generated rather than filmed, a course with forty lessons can have forty intros that look consistent without forty recording sessions, and each one can be re-cut by editing the script instead of booking a camera.
Sales follow-up (56s)
A fifty-six second personalised follow-up: greet the prospect by name, restate the problem, point at the next step. Sales teams rarely record these because the effort outweighs one email. Here the script changes per prospect and the face does not, so a hundred follow-ups cost a hundred credits and an afternoon of copywriting rather than a studio day.
Support answer (56s)
A fifty-six second answer to one recurring question, filmed once and embedded in the help centre. Support writing is already scripted, so the only missing piece is a face; this turns a paragraph of documentation into something a user will actually watch. When the answer changes, you regenerate the clip instead of re-recording a presenter.
Long-form narration (70s)
Seventy seconds, which is past the single-clip ceiling, so the audio was split at pauses, generated in parts and stitched back together. The picture is cut at those pauses and the soundtrack is the original recording, so the voice never drifts. This is the shape to test when your script runs long: watch the seams and judge whether the joins are visible.
How each clip was made
Every video started as one AI-generated portrait and one voice track, either uploaded or produced with the built-in TTS. Generation runs on a single H3 workflow call per thirty seconds of audio; anything longer is split at natural pauses, generated in parts, then rejoined with the original soundtrack laid back over the picture. Nothing was colour graded, stabilised or retimed afterwards.
What to judge before you buy
Watch the mouth at the start and end of each sentence, then watch the hands and the background. Lip sync is the part that holds up best; hands and fine background detail are where a generated portrait shows its seams. If your use case puts a person in front of a busy street, generate a portrait with a simpler background first and compare.
Why the examples are vertical
Nine by sixteen is the default here because short video is watched on a phone, and a vertical portrait fills that frame without letterboxing. Landscape and square are both available on the create page, and a portrait generated in one ratio will not fill another cleanly, so pick the ratio before you generate the face rather than after.