LTX 2.5 API

LTX 2.5 Text to video generates one continuous take of 5 to 20 whole seconds from a written brief, in 16:9 or 9:16, at 720p, 1080p, 2K or 4K. Only prompt is required; duration_seconds, aspect_ratio and resolution default to 5 seconds, 16:9 and 720p, and seed is optional. Every clip comes back with a generated soundtrack. Reach for this endpoint when nothing visual has been approved yet and the setting, action and camera all have to come from words. Requests run as asynchronous tasks: create the task, then poll it or take the callback.

Choose a LTX 2.5 mode by starting material

Up to twenty seconds in a single take

duration_seconds takes whole seconds from 5 to 20; the floor is 5, not 3. Twenty seconds is enough for an entrance, a hold and an exit inside one generated shot, so continuity across those beats is handled inside the generation instead of in your edit. The family has no extend task, which means length is set when you submit: a five-second draft and a twenty-second final are two separate generations, not one clip lengthened.

Four resolutions behind one field

720p, 1080p, 2k and 4k all come from the same request shape, and 2k returns a frame about 2560 px wide. Blocking and camera path can already be judged at 720p. What the higher tiers add is fabric weave, skin texture and specular detail, which is why the practical path is a 720p draft followed by a single high-resolution run with the seed unchanged.

Sound generated with the picture

The clip includes an audio track even though the endpoint has no audio field, so there is no switch to turn it off. Treat sound as part of the brief: the opera example ends with "orchestra tuning drifting up from the pit". When the clip goes under your own music and foley, mute or replace the track in the edit.

A long move with no cuts

The arena-gate-to-cages example runs at the 20-second maximum in 1080p and 16:9. Its brief describes one Steadicam path from an arena gate, through a packed terrace and a torch-lit tunnel, out to backstage cages, and says "No cuts" outright. Shots like this depend on the camera path and the set holding together at the same time, which is what a single long take is for.

Drafts that become 4K plates

Draft the take at 720p, keep the seed, then run the same prompt once at 4K; the resolution field is the only change. A 4K render carries enough detail to survive a grade, a reframe and a slow push added later in the edit. The copper-tub-candlelight example was rendered at 2k; its brief asks for steam, candle flame and fine surface texture, the kind of detail the higher tiers are there for.

Vertical, first-person and illustrated shots

In 9:16 at twenty seconds, a runway pass fits entrance, turn and exit in one generation. The bedside-pov brief opens by stating that the camera is a person lying against the headboard, which turns an observed scene into one happening to the viewer. There is no style field on this family, so the ramen-stall example sets its cel-shaded register entirely in the wording: "Anime cel-shaded animation, hand-painted backgrounds".

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "lightricks/ltx-2.5/text-to-video",
  "input": {
    "prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "duration_seconds": 5
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.