Puppet

What Puppet does

Give Puppet a character image and an audio clip and it renders a video of that character speaking. Motion is not decoration on a timer: one beat track is derived from the audio and every channel (brows, head, gaze, breathing, blinks) is a function of it, so the head dips and the brows lift on the same stressed syllable rather than drifting independently. Illustrated characters take a completely different path from photographic ones, because a model trained on real faces paints a photoreal mouth onto cel artwork; for drawn characters the mouth is constructed from colours sampled out of the character’s own pixels instead.

• Separate rendering paths for photographic and illustrated characters
• Blinks, gaze, head motion and breathing all driven by one audio-derived beat track
• Runs on Jalaran’s own GPU, and your likeness and voice never go to a third-party AI service
• A job completes only when the output file itself verifies

How Puppet works

1. Upload a character image

   A face scan runs first and fails fast if there is no usable face, rather than letting you wait minutes for a job that cannot work.

2. Choose photographic or illustrated

   You set this; it is never guessed. A wrong guess is invisible until playback, which is the worst time to find out.

3. Add the audio

   Audio from Voice carries a phoneme timeline and lip-syncs exactly. Any other clip falls back to estimating mouth shapes from the waveform.

4. Render on Jalaran’s GPU

   The job runs on Jalaran’s own GPU. Its estimate is based on measured throughput for the path you chose, not a generic guess.

5. Verify and download

   A job is only marked complete after the finished file is re-inspected: duration, frame count and decodability all have to match.

What Puppet does not do

Puppet renders on Jalaran’s own GPU, one job at a time, so a render may wait its turn; there is no cloud fallback. Audio length is capped, and the cap differs by path because the cost does. Generated media holds a person’s likeness and voice, so everything is purged on a retention schedule rather than kept. It animates one still character; it is not a video editor.

Common questions

Does my character image or voice get uploaded anywhere?

Only to your own Jalaran account: the files are kept in its private storage and deleted automatically after 30 days. Rendering happens on Jalaran’s own GPU, through a worker with no credentials and no database access, and no third-party AI service sees them. The files are stored by Supabase in Singapore, as the Privacy Policy says.

Does it work with anime or illustrated characters?

Yes, and that path is built for them specifically. A model trained on real faces generates a photoreal mouth on cel artwork, so for illustrated characters the mouth is drawn from colours sampled out of the character’s own pixels instead.

How do I know the video is not silently truncated?

Because a job is not marked complete until the output file is re-inspected: it has to exist, decode, and match the audio duration and the composited frame count. A wrongly-failed job is honest and retryable; a wrongly-completed one is indistinguishable from a good one.