LTX 2.5 Review: Long Takes With Sound, Uneven Faces

We don't publish our content-ladder runs; the gallery below is the everyday register, taken from the published, safe-for-work examples on the LTX 2.5 model page.

Sunset to firelight, in one request. Two people on a beach at sunset land, by the closing frame, wrapped in a blanket by a bonfire at night — the prompt only has to describe the passage between the two stills, not the destination.

Candlelight and rising steam. From an opening frame in a candlelit bathroom, she sinks into steaming water, hands dragging through soaked hair — the still supplies the room, the prompt supplies only what changes.

A backwards fall into a pool. She tips backward off a tiled ledge, bubbles rising past her as the camera sinks with her and caustics keep rippling on the columns — the kind of physics beat that usually breaks an animated still.

A twenty-second unbroken crowd crossing. One continuous Steadicam-style move from an arena gate through a packed crowd, written entirely from text with no cuts, at the endpoint's full 20-second length.

A marble staircase, descended in one take. A woman in an emerald silk gown descends an empty opera house's staircase as chandelier light catches the fabric — a full arc of descent, fabric motion and a crane move from a single paragraph.

A cel-shaded ramen stall in the rain. Anime-style, hand-painted backgrounds: a woman ducks under a noren curtain, shakes off an umbrella and sits — style carried entirely by wording, since there's no separate style field.

Order the sentence as subject, action, camera, light, then any constraint, and change one variable at a time when troubleshooting a bad take. On image-to-video, let the opening (and closing) frame carry identity and wardrobe, and use the prompt for motion only.

1. A pinned ending, described as passage (image-to-video, closing frame supplied).

Cast: one man and one woman. Begin on the two of them on the beach at sunset and end on the supplied final frame: the same two at night wrapped in one blanket by a bonfire, sparks rising. Close: wet swimwear still clinging at sunset, then opening onto firelight.

Run this prompt on LTX 2.5 →

2. A long take, written as one unbroken move (text-to-video).

Cast: a mixed crowd of men and women of different ages and skin tones. The camera stays at chest height in the crush of bodies the whole way through. One continuous 20-second Steadicam take: from the arena gate, through a packed crowd, no cuts.

Run this prompt on LTX 2.5 →

Weak: describing a fast, complex physical action (a canyon flythrough, a wok toss over flame) with heavy camera movement and fine detail in the same sentence — that combination is where artifacts were most severe. Strong: slow it down, or split a complicated action into fewer simultaneous camera and subject changes.

Between 16 and 24 September 2026 we ran our five-rung content ladder on both LTX 2.5 tasks. Image-to-video rendered L1 Suggestive, L2 Mature themes, L3 Strong and L4 Intense clean on every run, and the L5 Extreme restraint/impact probe rendered clean as well; we did not test the L5 Extreme gore probe on this family. Text-to-video matched that pattern except for one of its L4 Intense runs, which came back softened. Neither task carries a content filter of our own; the page label is Uncensored.

A note on sound and dialogue. Our audio tests on text-to-video found the results uneven rather than uniformly good or bad: a cafe dialogue scene delivered a natural laugh and mostly on-time lines with one inaccurate lip-sync moment; an emotional monologue landed the emotional beat but with a slightly artificial-looking face; a close-up "I missed you" line had good lip sync alongside odd mouth-shape artifacts partway through; a street-musician scene produced audible strumming with somewhat repetitive motion; a glass-shattering sound effect prompt rendered the wrong object (a tumbler rather than a wine glass) with an unclear shatter. Review the audio track on every clip before you publish it — there's no field to disable sound generation if a scene doesn't call for it.

Not tested yet: the L5 Extreme gore probe on either task; whether a supplied closing frame reduces the face-drift issue reported by independent testers; two-men scenes on the ladder; whether slowing down a complex-motion prompt avoids the colour-artifact failures we saw on canyon and wok scenes.

Adults only, in every prompt, still and output. Never depict a real, identifiable person without their consent, and hold the rights to every photo you upload as a frame.

Reaction to LTX 2.5 is split along almost exactly the line our own testing found. On the positive side, one Hugging Face user testing on consumer hardware wrote that "The generation speed and visual quality have been very impressive on my RTX 5070 Ti", particularly for landscapes and product shots.

The complaints cluster around consistency. A discussion titled "The character consistency in i2v is a disaster." reported the same reference photo rendering as a visibly different person between LTX 2.3 and LTX 2.5. A second thread, titled "Don't Use LTX 2.5 For Music Videos! Stick To LTX 2.3", pointed at the newer version's text encoder as the source of weaker lip sync for singing and dialogue.

Our own runs land on both sides of that split: the everyday jazz-club and staircase scenes we ran held together well, while the face-drift and lip-sync complaints from the community match what we saw in our own audio tests and side-by-side comparisons.

• Need a face that holds and don't mind giving up native audio? LTX 2.3 Spicy runs the older, more consistency-tested LTX generation with built-in LoRA presets, from $0.019 a second.
• Want an open-weight base with a deep local LoRA ecosystem instead? Wan 2.2 trades the long single-take format for a fixed five-second clip at a flat $0.24.
• Longer takes with multi-shot direction? Wan 3.0 runs up to 30 seconds, priced by the second from $0.045 at 480p.
• A different open-weight rival with its own community following? MiniMax H3 is the other family independent testers compare LTX against most often.