LTX architecture
and shot composition.

LTX-Video 0.9.x (2B and 13B variants, Nov 2024 through May 2025) was not the first open-weight video model — AnimateDiff, Stable Video Diffusion, and CogVideoX predate it. Its genuine first: faster-than-realtime DiT video generation on a single GPU — a 5-second clip at 768×512 @ 24fps in under 30 seconds on a 4090. LTX-2 (Oct 2025) added native synchronized audio, 4K output, and kept the weights open. Architecture: a DiT-style transformer trained jointly on text-to-video and image-to-video. ComfyUI is the de facto local orchestrator for open-weight video — Wan 2.x, Hunyuan, and LTX-2 all run through it.

Modern video models reward thinking in shots rather than seconds. A 5-second clip can be one slow pan, three cuts, or a single dynamic action — you control this through prompt language and motion-guidance inputs. Macalinao Studio practice: write the shot in script form ("opens on wide establishing shot, slow push-in over 3 seconds, then cut to close-up") before generating.

LTX-Video and Sora both accept motion specifiers: "static camera", "slow dolly forward", "orbit left", "crash zoom". Some systems also accept trajectory inputs — a bezier curve through 3D space that the virtual camera follows. This is the bridge between AI generation and traditional film grammar.

Ship today: native audio+video single shots — Veo 3, Sora 2, and LTX-2 all generate synchronized sound and dialogue lip-sync in one pass (native audio is LTX-2's headline feature) — plus controllable camera moves and stylized aesthetics. Still hard: multi-minute, multi-character narrative with identity and continuity holding across many cuts. The studio strategy: design around the strengths (short controlled shots, native or voiceover audio), stitch longer narratives in the edit.