Livepeer Agent · Fal Expansion · video · audio-driven

Audio-First Scene · LTX 2.5

Two things this family does that nothing else here does: it generates video from an audio track, and it puts a real draft→final ladder inside one model — the same prompt runs on fast and pro, so promoting a take costs a re-render, not a re-write.

Format
Video · native audio
Caps
ltx-25-{t2v,i2v,a2v}-{fast,pro}
Cost
$0.0945/s fast
$0.126/s pro (720p)
Persona
Creator · Music

The ladder, side by side

Same prompt, same seed of an idea, one tier apart. Both rendered for this page.

fast · ltx-25-t2v-fast · 6s @ 720p · rendered in 28.1s · $0.57

pro · ltx-25-t2v-pro · 6s @ 720p · rendered in 32.4s · $0.76

Generated from a voice track

ltx-25-a2v-pro — the audio came first; the motion is generated to it, not synced to it afterwards. Billed per second of input audio.

Why this works: the usual order is generate-then-mux — make a cut, lay the track under it. That is far cheaper and usually right. Audio-to-video is for when the audio came first and the motion has to land with it: a hook, a VO line, a beat. Getting that from a mux means cutting to fit, and it shows.

What we measured

All six endpoints baselined on the production MCP before shipping — 16/16.

CapabilityProbeResultWall clock
ltx-25-t2v-fast6s · 720p · n=33/3p50 28.1s · p95 30.3s
ltx-25-t2v-pro6s · 720p · n=33/3p50 32.4s · p95 32.6s
ltx-25-i2v-fast6s · 720p · n=33/3p50 33.0s · p95 42.9s
ltx-25-i2v-pro6s · 720p · n=22/2p50 33.1s · p95 35.9s
ltx-25-a2v-fast~6s voice · n=33/3p50 32.7s · p95 39.4s
ltx-25-a2v-pro~6s voice · n=22/2p50 39.7s · p95 41.0s
Duration barely moves it: the same fast t2v at 20s took 42.9s — ~22s fixed + ~1s per second of video.

For comparison, Seedance 2.5 needs ~190s for a 4-second take. At half a minute a render, showing someone three variations here costs less wall-clock than one draft there — which is what makes the ladder usable rather than theoretical.

The honest limits

Open playbook → ← All playbooks All showcases →