LiveWan

streaming text-to-video · Wan2.1-1.3B student · step 3000
connecting
Loading the model
This takes about 90 seconds on first start.
block—
speed—
streamed—
latent frames—
K/V cache—
buffer—
Now conditioning on
—

What you are looking at

A 1.3 B student, distilled from a Wan2.1-14B teacher, generating video continuously instead of as a fixed clip. Every 750 ms it produces another block of 12 frames conditioned on a rolling K/V cache of what it already made — so the picture never restarts.

STEER · continues
Swaps the text, keeps the cache. The scene morphs in place. Best between related scenes.
SCENE · cuts
Clears the cache and reopens from a cached world. Use it when you want a genuinely different picture.

1 · Scene ?A scene is one of four cached worlds — 21 latent frames the base model generated once, so a stream can open instantly instead of waiting ~60 s. Picking one cuts: the stream restarts and the K/V cache is cleared.

Pick a starting world. This cuts to a new stream.

2 · Steer ?Steering swaps the text conditioning but keeps the K/V cache, so the scene carries on and morphs instead of cutting. Nearby scenes blend beautifully; a violent jump (day → night) tends to smear.

Change the text mid-stream. The scene continues.

instant
Advanced

Defaults match the published command. Steps above 2 cost proportionally more time without fixing hard steers; latent norm 0 lets long streams blow out.

3 · Look ?Display-only post-processing, applied in your browser as frames arrive. It costs the generator nothing, so the stream stays at 1× — but it also does not change what the model produced. Useful against the softness long streams drift into.

Post-processing on the picture. Costs no GPU time.

0
1.00
1.00
1.00