Reactor LingBot World 2 Segment

LingBot World 2 is LingBot with better camera manners. Same "one image, a world you move through" idea, but the move lanes gain preset poses - orbit, dolly, crane - so you're writing camera language instead of nudging direction and hoping.

What it's for

The local story on this model is worth knowing, because it explains why the hosted version is the one most people will actually touch. LingBot-World 2.0 is a 14B causal model built on Wan 2.2, real-time at 720p, holding coherence for up to an hour - and it's CC-BY-NC-SA, non-commercial and share-alike, with no ComfyUI path on release. Through Reactor you skip all of that: one image in, a driveable world out, no license reading required.

You'd pick this node over plain LingBot when the camera work is the point. A slow orbit that reveals the horizon, a crane up out of a valley, a dolly in on a doorway - the presets make those expressible as blocks on a timeline instead of prose in a prompt.

How it works

One image on the first segment, required, read at the start and held. The model chunks at 24 frames with a 17-frame first chunk, so frames steps by 24 and both ends round to the nearest chunk, and the plugin writes the output at 48fps. Compared with LingBot's ceiling on a single scene, World 2's is effectively unlimited here - this is the model in the pack whose whole pitch is long, continuous generation.

The moves widget is the star. On top of LingBot's direction lanes you get a camera-pose lane offering orbit_left, orbit_right, dolly_in, dolly_out, crane_up and crane_down, plus longitudinal (forward/back) and lateral (strafe left/right) movement, look up/down/left/right, and a rotation speed. Each move is a block with a start frame and a length measured from its own segment's start, and a move ends when its segment does - continuing a camera move across a boundary means declaring it again on the next segment.

The inputs that matter

• prompt - what the world is and what's in it. Reactor links a prompt guide from the example notes; world-model prompting is not the same as clip prompting, so read it.
• image (optional in the schema, required in practice) - the world's opening picture. First segment only.
• frames - default 120, stepping by 24.
• moves - camera pose, movement, look and speed, as blocks on the lane.
• sequence - the earlier segments; empty on the first.
• REACTOR_SEQUENCE (output) - onward, then Reactor Render, or the visualizer to check your block alignment.

Install

Manager search comfyui-reactor-api ("Reactor API: Realtime Video"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/reactor-team/comfyui-reactor-api
pip install -r comfyui-reactor-api/requirements.txt

One dependency (reactor-sdk>=1.6), nothing to download. Set REACTOR_API_KEY (environment, or config.ini under [API]), restart, and you'll find the pack under Reactor. The keyboard-driven world example in the README runs through Reactor Realtime, which is where this model is most fun - W/A/S/D and the arrow keys, live.

Where people get burned

Camera moves are per-segment. The most common wasted render: a beautiful dolly on segment one, then nothing on segment two, because a move ends with its segment.

Moves can't be shorter than a chunk. Anything under 24 frames gets rounded up or flagged red on the visualizer.

Presets are relative, not absolute. orbit_right turns a fixed amount rather than to a coordinate. Chaining two of them turns twice as far; if you wanted one smooth 180, use the rotation speed lane instead of stacking poses.

The world doesn't remember. Per the KB's honest accounting of this family, scene state doesn't persist - revisit an area and it regenerates rather than reusing what was there, and there's no real collision physics. Script your moves through the scene rather than going back to it.