Reactor LongLive-2.0 Segment

Here's the thing that surprises people about this node: there's no image socket. LongLive-2.0 generates from text, full stop. No reference picture, no first frame, no img2img foothold. And of all the segment nodes in this pack, that makes it the one most about structure - how a story moves between shots.

What it's for

LongLive-2.0 is the "a story across shots and cuts" model in this pack, and the workflow example that ships with it is nine segments of exactly that. You write a prompt per segment, decide whether each one cuts or blends, and the model plays them as one continuous piece of video. It's the closest thing here to editing rather than generating: you're building a sequence of beats and the model handles continuity.

Reach for it for B-roll chains, for a scene that needs to hold across a cut, for anything where the shape of the timeline matters more than auditioning a single clipping. If you want a specific person or place, you're on the wrong model - nothing in this node will condition on a picture.

How it works

LongLive is a chunked model: 32 frames per chunk with a shorter 29-frame first chunk, at 24fps. Your frames value gets rounded to the nearest chunk edge, which is why the field steps by 32, and why the timeline's boundaries may not land exactly where you typed them. kind is the other half of the scheduling: shot blends softly from the previous segment, cut starts a fresh scene, and a cut resets the chunk grid underneath it.

There's also a hard ceiling per scene - the model restarts after a set number of chunks - so a scene can't run forever no matter how many frames you ask for. Plan a cut before you hit it rather than discovering it in the output.

The inputs that matter

• prompt - this segment's shot. Read the model's prompt guide; the pack links it from the example notes and it applies directly.
• frames - default 128, stepping by 32. Rounded to the nearest chunk at both ends.
• kind - shot or cut. The only structural control this node has, and the one that does the most work.
• sequence - the segments before this one; leave empty on the first.
• REACTOR_SEQUENCE (output) - the next segment, and finally Reactor Render (or Reactor Sequence Visualizer first, which is where you'll notice the chunk rounding).

Install

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

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

reactor-sdk>=1.6 is the entire dependency list and there are no model files - generation happens on Reactor's hardware. Set REACTOR_API_KEY (environment, or [API] in config.ini), restart, and the nodes are under Reactor.

Where people get burned

You can't sneak an image in. There's no socket, and the sequence compiler rejects one outright for this model. If a node upstream produces an IMAGE, that's for a different segment node.

Progress looks wrong on long timelines. LongLive counts emitted frames across the whole session rather than per chunk, so a progress readout can look out of step with the chunk you think you're on. It's cosmetic; don't interrupt a render over it.

Chunk rounding compounds. Each segment rounds independently, so six segments of 128 frames won't total what you multiplied out. Check the total on the visualizer before a long render - that's what it's for.

Nine segments is a lot of prompting. Since there's no image conditioning to lean on, continuity has to come from the text: name the same subject and place explicitly in every segment's prompt, or the model will quietly recast your scene at each cut.