🎬 LTX-2.5 · Text-to-Video Builder

[ diffusers weights ] [ native weights ] [ official demo ]

LTX-2.5 Distilled (22B) — video with synchronized audio, in 8 + 3 steps, no CFG. Runs end-to-end on upstream diffusers over ZeroGPU.

Prompt

Aspect ratio

Sets width/height on the two-stage /64 grid.

Landscape 16:9 Portrait 9:16 Square 1:1

Duration (seconds, 24 fps)

Snapped to the VAE's frame grid (8k+1 frames).

1 5

Auto length — let the model choose (≤15s)

Uses LTX-2.5's duration head to pick the length from the prompt, overriding the slider above. Off by default: the length is only known once the run is already on the GPU, so an auto run has to reserve ZeroGPU time for the full 15s worst case whatever it ends up generating.

Width (final)

512 1536

Height (final)

512 1536

Steps (fixed: 8 in stage 1 + 3 in stage 2)

Seed

0 2147483647

Randomize seed

Result (with audio)

Seed used

Length generated

What the duration head chose, when auto length is on.

Examples — text-to-video, with audio

Runs on ZeroGPU (xlarge slice). Each generation uses your free daily Hugging Face GPU quota — larger resolutions and auto length reserve (and therefore spend) more of it.