Realtime Videos · Models
LongLive-1.3B
Autoregressive text-to-video that generates in blocks and lets you switch the prompt mid-stream.
Overview
LongLive generates video one block at a time, each block conditioned on the ones before it. That is what makes a mid-stream prompt change continue the scene instead of restarting it.
| Property | Value |
|---|---|
| Model name | longlive-1.3b |
| Parameters | 1.3 B |
| Output | 480×832, 16 fps, one video track (main_video), no audio |
| Block | 12 frames = 750 ms of video (the first block is 9 frames) |
| Commands | set_prompt, set_paused, set_seed (see Schema) |
| Max session length | 240 s of video |
| Concurrency | One session per GPU pod |
Behavior to design around
- Prompt changes land at block boundaries. Expect up to roughly one block (750 ms) plus generation time before a change is visible.
- A prompt-switch block is slower. The model rebuilds its context from the recent frames, so the block that applies a switch takes longer than a normal one.
- Seed is read once.
set_seedapplies when the stream starts; changing it later has no effect. - A slow GPU shows as a held frame. If a block takes longer than 750 ms to produce, playback holds the last frame; it does not error.
Performance
We have not yet measured step latency, capacity, or time to first frame for this model on production hardware, so we publish none. The model's authors report about 20.7 frames per second on an H100 for generation alone, excluding video decode. Treat that as an upper bound on an unmeasured stack, not as what you will get. We will publish our own measurements, with the exact configuration that produced them, when we have them.
Tracks
| Name | Kind | Direction |
|---|---|---|
main_video | video | out (pod to client) |
Commands
| Command | Payload | Default | Notes |
|---|---|---|---|
set_prompt | prompt: string, at most 512 chars | A calm lake at sunrise, mist rising off the water, cinematic. | Applied at the next block boundary by re-building context. Marked for moderation. |
set_paused | paused: boolean | false | Holds generation on the last frame. The session keeps its GPU slot and keeps billing. |
set_seed | seed: integer, 0 to 2147483647 | 0 | Read when the stream starts. Later changes have no effect. |
Example
{
"model": "longlive-1.3b",
"max_duration_s": 120,
"starting_input": { "state": { "prompt": "a red fox running through fresh snow", "seed": 7 } }
}The authoritative, machine-readable version is GET /v1/realtime/sessions/{id}/schema on a live session.