AI repair guidance video: Griffin's idea, Sume's step clips

Live guidance and fixed guidance

Tavus's Griffin page, read 2026-10-03, lists technical troubleshooting and repair guidance among the settings for its full-duplex video model. The picture is a person showing a broken part to a camera and a model that sees it and talks them through.

That depends on a model that is open to customers. Tavus says Griffin-Lite is offered only to select trusted testers. Meanwhile, most repair help is fixed: the same twelve steps for the same product, which you can render once.

A step clip on Sume

Each repair step becomes one avatar video job. The avatar video endpoint takes a script, an optional product_image, and a scene that is either a prompt or a photo. Use the product photo as the scene reference so the avatar stands next to the real part rather than an invented one.

Captions can be burned in with the captions object, so the steps read in a noisy garage. The slam, punch and tiktok-green styles are Latin-script; Korean speech needs a Hangul style or the request is rejected with caption_hangul_text_latin_style.

curl -X POST https://api.sume.com/v1/avatar-1.0/talking-video \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: repair-step-03" \
  -d '{
    "avatar_handle": "repair_host",
    "aspect_ratio": "9:16",
    "script": "Step three. Unplug the unit, then press the release tab and slide the filter out.",
    "product_image": "https://example.com/unit-photo.jpg",
    "captions": {"enabled": true, "style": "punch", "language": "auto"}
  }'

What to keep honest

• One job is 4 to 60 seconds, so one clip per step keeps each under the cap and lets you re-render a single step.
• The avatar says what you wrote. It does not see the customer's unit, so write the script around what is true for every unit and point to a model number.
• The avatar scene is built from a prompt or photo, and the product image is a reference, not a diagram. For wiring or part numbers, use the manufacturer's drawing alongside.
• Safety steps (unplug, depressurise, wear gloves) belong in the script and in the captions.

Keep the set maintainable

Name each job after the step, keep the scripts in a file, and store the job ids next to them. When a part changes, re-render only the affected step. Because each job takes an Idempotency-Key, give each step a stable key such as repair-step-03-v2 and bump the suffix when the script changes, so a retry of an old request cannot return a stale clip.

Preview the first frame of the opening step before you render the rest, so the avatar and product placement are settled once for the whole set.