LTX-2.5: The Audio and Video Model Inside tendre.AI

LTX-2.5 is the audio-and-video generation model that powers video inside tendre.AI. It belongs to the LTX family of video diffusion models and its defining trait is that it generates picture and sound together, in one model, rather than producing a silent clip and leaving audio for a separate tool. This article is a plain-language technical look at how it works and how tendre.AI integrates it.

Why a joint audio-video model matters

The old way to get a clip with sound was a pipeline: a video model makes the frames, an audio model makes a track, and you align them by hand. The problem is alignment. Sound that was generated without "seeing" the motion never lands exactly: a footstep half a beat late, a voice that does not match the mouth, ambience that ignores the scene.

LTX-2.5 collapses that pipeline. Because the same model produces the frames and the audio, the two are coherent by construction: the soundtrack is conditioned on the same content as the picture, so motion and sound are synchronized from the first generation, not patched afterwards.

The architecture, in plain terms

LTX-2.5 is a diffusion transformer (DiT). Two ideas worth understanding:

• Diffusion means the model starts from noise and denoises step by step toward a clip that matches your prompt. It is the same principle behind modern image models like SDXL, extended to time.
• Transformer means it attends over the whole sequence (across frames, and across the audio stream) instead of treating each frame independently. That global view is what keeps motion stable and the audio locked to the action over the length of the clip.

Working over the full clip at once, rather than frame by frame, is the core reason the output stays coherent: objects keep their shape, the camera move stays smooth, and the sound tracks the picture.

Text-to-video and image-to-video

LTX-2.5 supports two entry points, and tendre.AI uses both:

• Text-to-video: describe the shot, get a clip with sound.
• Image-to-video: start from a still you already generated locally in tendre.AI and animate it. The first frame is your image, so the character and style you locked in (with a LoRA or a fixed seed) carry straight into the video.

Image-to-video is what makes the "one tool for image and video" workflow real: the picture you love becomes the opening frame of the clip, same face, same look.

Resolution: 1080p for iteration, 4K for finals

The same model targets multiple resolutions. In practice that gives a clean workflow:

• 1080p (Full HD) for iteration: fast enough to try a prompt, hear the result, adjust, and run it again.
• 4K (Ultra HD) for final renders: four times the pixels, for big screens or room to crop and stabilize in post.

You draft at 1080p, lock the shot (motion, framing, audio), then finish the keeper in 4K, without switching engines between draft and delivery.

Efficiency is the point

The LTX line is known for being fast for its quality. That efficiency is not a vanity metric: it is what makes quick 1080p drafts and on-demand 4K finals practical instead of overnight jobs. A model that is efficient enough to iterate with changes how you work, you explore more takes because each one is cheap in time.

What version 2.5 brought

Lightricks released LTX-2.5 with open weights on 11 August 2026. Four of its changes matter for anyone actually generating clips.

Native multi-shot scenes. The model can render connected shots, wide then medium then close-up, in a single generation, holding the character, the setting, the lighting and the voice steady across the cuts. Before, a sequence meant rendering each shot separately and hoping the face and the light matched. This is the change that turns a clip into a scene.

A new diffusion video decoder. The most visible gain on ordinary footage: fewer artifacts in high-motion scenes, precisely where earlier versions smeared. A head turning quickly, a fast pan, hair moving, all hold together better, while the high compression ratio the LTX family relies on is preserved.

A better text encoder. Prompt handling moved to a fine-tuned Gemma 4 12B encoder paired with a dedicated prompt enhancer. In practice it reads complex, multi-subject instructions from shorter prompts, so you spend less time over-specifying.

Automatic duration. Clip length is inferred from the action you describe rather than set by hand in frames, and native 4K HDR is now part of the pipeline instead of a post-process.

For scale: on datacenter hardware, LTX-2.5 renders 10 seconds of 720p in under 7 seconds. That number is what makes cloud rendering billed per clip viable at all.

These are capabilities of the engine. They land in tendre.AI progressively, release after release. The changelog is the reference for what is live in the version you are running.

How tendre.AI integrates LTX-2.5

tendre.AI applies its usual rule: local first, cloud only for the heavy lifting you choose.

• Images stay 100% local. Every still is generated on your own GPU, nothing uploaded.
• LTX-2.5 video runs on a cloud GPU, on demand. Synchronized audio and especially 4K are compute-heavy, so they run on a remote GPU and are billed per clip in credits. It is opt-in: if you only generate images, nothing about your private local workflow changes.
• Same characters across both. Because video starts from your local stills, the identity you built carries into the clip.

Migration note: tendre.AI is actively rolling LTX-2.5 into the app. Video with sound, 1080p iteration and 4K finishing land progressively as the migration completes. The local image workflow is unaffected.

The content boundary still applies

LTX-2.5 does not change tendre.AI's firm rule. Everything generated is 100% synthetic: no real person is depicted, and every subject is unmistakably an adult. The model is a tool for fictional, adult, AI-generated content, nothing else.