HeyGen scene-scoped edits vs fixing one scene in a Sume avatar

HeyGen's Video Agent now accepts a scene-scoped edit plan that leaves untouched scenes alone. Sume's avatar video API has no equivalent edit call: once a video is rendered, changing the script means a new render, and the cheap place to catch mistakes is the preview stage before it.

This post lays out what HeyGen documents, what Sume documents, and the workflow that follows for a script that keeps changing.

What did HeyGen add for editing one scene?

The September 2026 changelog entry says POST /v3/video-agents/{session_id} accepts an edit_plan: one natural-language change per scene, applied together in a single turn. Untouched scenes keep their script and visuals. You need scene IDs from GET /v3/videos/{video_id}/scenes, and a request can carry up to 50 items, with validation that the scenes are consistent.

That is an edit against an existing video session.

What can you change on Sume without starting over?

Sume splits the work in two. An avatar video preview generates only the first-frame stills, one per input scene when you use video_inputs. You approve the composition, then call generate-video on the preview id and Sume reuses the stills.

Two things are explicitly cheap to change after approval. quality can be overridden at generate-video (standard, plus or max), because preview stills are tier-independent. And POST /v1/avatar-video-previews/{id}/regenerate reuses the stored request and refreshes only the first-frame stills.

What forces a new preview?

The docs list the structural fields: script, video_inputs, avatar_handle, scene and aspect_ratio. Change any of them and you create a new preview. There is no documented call that patches one scene of a finished avatar video and keeps the others.

How do you work when only one sentence changes?

Plan for it before you render. Keep each spoken beat in its own short job when the beats are likely to change, and join the finished clips later with Timeline 1.0, which handles sequencing and transitions. A fix then re-renders one short clip instead of the whole video.

The trade-off is real. Separate avatar videos are separate jobs, each with its own 4-60 second window and its own cost, and each is rendered on its own, so the framing of a shared scene is not guaranteed to match between them. Within a single multi-scene video, Sume resolves one avatar and one shared scene.

Is the preview step worth it for edits?

Yes when the risk is composition: framing, product placement, background. No when the risk is wording, because stills do not show what the avatar says. Review wording in your own tooling first, then preview, then render.

Captions are stored on preview create and burned in only at generate-video; preview stills are never captioned. For the related question of what happens after approval, read change the script after approving a preview, and for the whole-render angle fix one sentence in an AI avatar video.