Playable AI worlds vs pre-rendered branches on Sume

PixVerse R2, announced September 22, 2026, is a real-time world model that supports WASD exploration and branching narratives. Sume has no live world session. If you need a branching story rather than free exploration, pre-render each branch as a clip with start and end frames, then assemble the paths in Timeline.

What PixVerse announced

From the PixVerse R2 post, read on 2026-10-03:

What Sume does instead

Sume video jobs are asynchronous: submit, receive a job id, poll, download. That is the opposite of a live session, so interactive movement is out of scope. What it can do is produce the fixed set of clips a branching video needs.

Branch with frames

POST /v1/videos accepts frame_images with frame_type of first_frame or last_frame on models whose supported_frame_images lists them, such as seedance-2 in the docs example. Use a shared still as the end frame of the intro and the start frame of each branch, so every path joins the same visual moment.

To chain clips, extract a still from the end of a finished clip with POST /v1/video-frames (give at[] times), and use it as the next clip's first frame. Frames come back as durable media.sume.com images.

{
  "model": "seedance-2",
  "prompt": "The hero walks left through the door into the garden",
  "frame_images": [
    {
      "type": "image_url",
      "image_url": {"url": "https://example.com/shared-doorway.png"},
      "frame_type": "first_frame"
    }
  ],
  "resolution": "720p"
}

Assemble each path

Timeline 1.0 takes one audio spine plus ordered video[] slots and returns one MP4, with optional transitions between slots. Render one MP4 per path through the story: intro plus branch A, intro plus branch B. Imported inputs must already be media.sume.com artifacts.

The player, the choice buttons and the routing between files are yours to build. Sume produces the files, not the playback logic.

When to pick which

Choose a world model when the user steers in real time. Choose pre-rendered branches when you have a script with a handful of decisions, need identical output on every playback, and want per-job cost known in advance from usage.cost on each receipt.