LTX-2.5 native multishot: one prompt, or clips joined on Sume

LTX-2.5's model card on Hugging Face (read 2026-10-02) says the release introduces native multishot generation: connected scenes in a single pass, keeping character identity, environment and visual style across cuts. Sume does not list an LTX model, so the Sume route to a multi-shot video is to generate each shot as its own clip and assemble them with Timeline 1.0.

What the model card says

The Lightricks/LTX-2.5 card (read 2026-10-02) lists the items below. Other coverage dates the launch to Aug 11, 2026; this page relies on the card itself for the specs.

One pass versus one shot at a time

A single multishot generation has one advantage: the model sees the whole sequence, so a character who appears in shots one and three is more likely to match. The cost is control. If shot two is wrong you regenerate the sequence, or use whatever per-segment tools the pipeline offers.

Generating shot by shot reverses that trade. You can fix one clip without touching the others, but you carry consistency yourself, with references, a fixed first frame, or a shared prompt block.

The Sume route

Sume's video API generates a clip per job. Limits vary by model: for example Wan 3.0 accepts 2 to 30 seconds and Seedance 2.5 4 to 30 seconds per the video docs. To join shots, Timeline 1.0 takes one audio spine plus ordered video[] slots and returns one MP4, handling sequencing and transitions.

Models that accept image references, such as those listed in the catalog, can help keep a character stable across separate shots. Sume has no single-pass multishot switch, and does not claim cross-shot consistency from the join itself.

Decision notes

Check the model card again before shipping; specs and license terms can change.

• Short sequence, one look, you control the GPU: LTX-2.5 multishot may suit.
• Many iterations on single shots, hosted: separate clips plus a timeline.
• Revenue above $10M: read the LTX license terms before you build on the weights.