HeyGen Video Agent edit_plan vs Sume preview regenerate

HeyGen's Video Agent now accepts an edit_plan on POST /v3/video-agents/{session_id}: one natural-language change per scene, applied together in a single turn, with untouched scenes keeping their script and visuals. The Sume docs describe no equivalent in-place scene edit. Its closest flow is to regenerate the preview stills, approve them, then start the final render.

HeyGen facts are from its September 2026 changelog; Sume facts are from Avatar video previews and Generate avatar video, read 2026-09-30.

How does edit_plan work?

You read scene ids and edit_version from GET /v3/videos/{video_id}/scenes, then send items with scene_id, text, scene_snapshot_video_id and edit_version. A scene outside the snapshot returns 400 invalid_parameter; a snapshot that changed since you read it returns 409 stale_edit_version. Up to 50 items per request.

What does Sume offer instead?

Avatar video previews generate the first-frame stage without the full talking-video render, so you can approve composition before spending on it. POST /v1/avatar-video-previews/:id/regenerate reuses the stored request and refreshes only the stills. POST /v1/avatar-video-previews/:id/generate-video then starts the render, reusing the preview first frame when available.

Which changes need a new preview on Sume?

The docs say structural fields (script, video_inputs, avatar_handle, scene, aspect_ratio) still require a new preview. Only the final render tier can change at the generate step. Note that regenerate refreshes stills for the whole stored request, not one chosen scene.

How should I plan edits on Sume?

Put each scene in its own entry of video_inputs, review scene_previews[], and start a new preview when the script changes. For the wider editing picture, see edit an AI avatar video.