Griffin-style follow-ups with rendered avatar clips and branching

You can approximate a guided avatar conversation today by rendering a small library of short clips and choosing the next clip in your own code. It is not live and it does not perceive the viewer, but it covers the common case of a tutorial, a rehearsal or a support flow with a known set of paths. Sume Avatar 1.0 renders each clip from a script, and your app owns the branching.

What Griffin-Lite does and what this replaces

Tavus says Griffin watches, listens and replies as video on a live call, and lists tutoring, rehearsing difficult conversations and collaborative problem-solving as uses. The same page says Griffin-Lite is a research preview for select trusted testers and is not available to customers at this time. A branching clip library replaces only the predictable half of those use cases.

Design the tree first

Write the paths as a table before you generate anything. Each row is a clip id, the line the avatar says, and the ids it can lead to. Keep each line short. Sume accepts scripts of 4 to 60 seconds, so a clip under 4 seconds needs a longer line or a silence beat.

Limit depth. Three levels with three choices each is 13 clips, and that is already a lot to review. Most real flows need fewer, because many branches can end at the same shared closing clip.

Render the library

Create one avatar, then reuse its handle for every clip so the person looks the same on every branch. Call POST /v1/avatar-1.0/talking-video once per clip. Use the clip id inside the idempotency key, for example tree-v1-clip-07, so a retry returns the original job instead of a second render.

For a larger tree, the Format bulk-run queue lets you leave up to 100 runs to a server-side queue (Bulk runs). Plain avatar jobs also work: submit them with a small concurrency window and store each job id against its clip id.

Pick the next clip in your app

Your front end shows buttons or a text field. Your logic maps the answer to a clip id and plays the stored media.sume.com URL. If you want free-text answers, classify them with any model you already use and map the label to a clip. Sume is not in that loop at play time, which keeps playback instant.

Add a catch-all clip that says what the avatar cannot help with and where to go next. A dead end is the most common failure in these trees.

Be clear with viewers

Say plainly that the avatar is AI-generated and that the answers are pre-recorded. When Griffin-Lite or another live option opens up to you, you can keep the same script tree as the fallback for when it is unavailable.