Open-weight audio-video foundation model

Official model guide plus a working online generator

LTX 2.5 AI Video Generator

Explore native multishot storytelling, synchronized audio and video, automatic duration, Diffusion Fidelity Rendering, and professional 4K HDR output from Lightricks' new open-weight foundation model. The LTX 2.5 AI video generator just below lets you create from a prompt, an image, a pair of frames or supported references.

Generate now See what it can do

Official LTX 2.5 AI Video Generator showcase of multishot and HDR output — Lightricks' launch footage, taken from the official LTX 2.5 page and compressed locally so it plays smoothly on the web.

Try LTX 2.5 online

The full MuseGen video workspace is built in here with LTX 2.5 already selected. You can still try other variants or change models and keep the rest of your setup.

You need an account to generate, and credits depend on the model, variant, length, resolution and other settings you choose. Credits for failed tasks come back automatically.

LTX 2.5 explained: what it is and when to use it

LTX 2.5 is the 22-billion-parameter audio-video foundation model Lightricks released with open weights in August 2026 (on the 11th). It is meant to be a base for production rather than a one-trick clip maker: you can run it on your own hardware, fine-tune the pretrained checkpoint, call the hosted API, or assemble specialised workflows on top using LoRA and IC-LoRA. This version reworks both what you can create and how it is rendered. Native multishot moves through linked wide, medium and close views while keeping the character, the location, the light, the voice and the look intact across each cut. Longer instructions survive thanks to a bespoke Gemma 4 12B text encoder plus a dedicated prompt enhancer, and the new diffusion video decoder with Diffusion Fidelity Rendering puts detail where the scene needs it rather than spreading one fixed compression budget over everything.

Parameters in an asymmetric dual-stream diffusion transformer: 22B
Native high-resolution output for finishing: 4K HDR
Longest Fast-tier clip at 1080p: 20s
First-stage schedule in the distilled model: 8 steps

Native multishot

An LTX 2.5 AI Video Generator that keeps one world intact across cuts

The headline upgrade is more than another resolution option. LTX 2.5 builds linked scenes in a single generation and holds on to the visual and sonic details that make them read as one sequence. The official clips below first show cleaner movement and closer instruction following, then the multishot payoff: the camera moves nearer or further and reframes, yet subject, setting, lighting, voice and style all stay in the same world.

What improved for continuity

Cleaner motion

Closer prompt adherence

“Build a linked sequence that goes from a wide establishing view to a shot of the character and then to a tight detail. Keep the identity, setting, light direction, visual style and voice the same across every cut, and let each action complete before the next shot starts.”

Production controls

Let the action set the length, then take the result into finishing

Automatic duration matches the action, not a preset

Before generating, a compatible duration head looks at the scene and estimates the time the action you described will take. Run locally, you can skip the frame count entirely or set a minimum and a maximum. A quick reaction stays quick, while a move with a setup, an action and a recovery gets enough frames to play out — avoiding the pacing problems you get when every idea is forced into five or ten seconds. Exact frame counts are still there when a production needs precise timing.

Native HDR and EXR leave headroom for the grade

The HDR workflow takes scene-linear EXR conditioning in sRGB/Rec.709, ACEScg or ACEScct and can export half-float EXR frames next to a 10-bit BT.2020/HLG HEVC master. Rather than baking a viewing look into an eight-bit intermediate, colourists and effects artists keep detail in the highlights and shadows for compositing, grading and delivery. The hosted model also offers 1080p, 1440p and 4K tiers for work that has to live beyond a social feed.

Diffusion Fidelity Rendering puts detail where the shot needs it

Older pipelines use one compression and reconstruction budget for the whole scene. LTX 2.5 brings in a new diffusion video decoder and a Diffusion Fidelity Rendering pipeline that can generate in-between keyframes, add a full-resolution detail pass and, if you want, a temporal refinement pass. Effort is distributed by scene complexity, so fast motion, faces, fine textures and readable screen content improve without every part of every frame costing the maximum.

LTX 2.3 vs LTX 2.5

LTX 2.3 still works with the same repository and plenty of existing adapters, but LTX 2.5 changes the foundation for new production work. Checkpoint components and LoRAs do not always carry over, so test your adapters before you move existing workflows to the new release.

Feature	LTX 2.3	LTX 2.5
Scene structure	Mostly a single continuous shot	Native linked multishot sequences
Video reconstruction	Convolutional VAE decoder	New diffusion video decoder
Text understanding	Gemma 3 text encoder	Custom Gemma 4 12B with a prompt enhancer
Duration	Frame count set by hand	Manual frames or automatic duration predicted from the prompt
Finishing	Mostly SDR output	Native EXR, ACES-family conditioning, RAW and 4K HDR
Open checkpoints	Dev, distilled, quantised	Separate dev/distilled components plus a raw pretrained foundation
Rendering	Fixed multi-stage upscaling	Diffusion Fidelity Rendering with detail and temporal passes

Three production jobs an open foundation makes easier

The official clips below show the less glamorous side of generative video: keeping a high-dynamic-range master, getting more usable takes before review, and adapting the model to a private domain. LTX 2.5 runs as a hosted service, but what really sets it apart is that the same foundation can move into an on-premises or fine-tuned workflow.

Finishing to cinema standards

Send generated shots through an EXR, ACES, DaVinci Wide Gamut or HLG workflow and keep the latitude professional colourists rely on.

Ads that demand fine detail

Rely on the rebuilt decoder and Diffusion Fidelity Rendering for packshots, faces, materials and signage that must hold up on a full-size screen.

Customising the model privately

Retrain the base checkpoint on your brand, a recurring character, a specialist subject or proprietary data, and keep it running on-premises.

Generate now

Making a video with LTX 2.5, step by step

Take an idea from first prompt to a configured LTX 2.5 render without leaving the model page.

Write the shot

Write a prompt covering the subject, action, setting, camera, look, pacing and audio. Got references? Upload them and say what each one is there for.

Set up LTX 2.5

Leave LTX 2.5 selected, pick the variant, mode, length, frame shape, resolution and audio options that suit the shot, then look at the credit cost displayed.

Render, check, reuse

Start the job, watch it progress beside the generator, check the finished clip, then reuse the same settings for a fresh attempt or download the file.

Explore the toolkit

Video Upscaler

Video to GIFFree

Image extractor

Instagram grid makerFree

View all tools

LTX 2.5 AI video generator: your questions answered

What is the LTX 2.5 AI Video Generator?

It is Lightricks' open-weight foundation model for audio and video, launched in August 2026. A 22B asymmetric dual-stream diffusion transformer produces footage with matching sound, and this version adds native multishot continuity, a diffusion video decoder, automatic duration, better prompt understanding, Diffusion Fidelity Rendering and professional HDR/RAW workflows.

Which generation modes are available in LTX 2.5?

LTX 2.5 handles text-to-video, image-to-video, native multishot sequences, audio-conditioned generation, keyframe interpolation, retake, extension and adapter-based transformations. The exact controls depend on the workflow you choose, but your prompt should cover action, camera, lighting, sound and shot changes in the order they happen.

How does LTX 2.5 differ from LTX 2.3?

The biggest changes are native linked multishot generation, a new diffusion video decoder, a custom Gemma 4 12B text encoder, a dedicated prompt enhancer, duration predicted from the prompt, Diffusion Fidelity Rendering, native EXR/HDR/RAW support, a stronger distilled model and a raw pretrained checkpoint for deeper adaptation. Plenty of LTX 2.3 LoRAs and IC-LoRAs may still work, but Lightricks advises testing adapters because compatibility is not guaranteed.

Does LTX 2.5 output 4K HDR video?

Yes. The official Fast and Pro tiers offer 1080p, 1440p and 4K at 24, 25, 48 or 50 fps. The native HDR workflow accepts EXR inputs in supported scene-linear or ACES-family colour spaces and outputs half-float EXR frames along with a 10-bit BT.2020/HLG HEVC master. Fast reaches 20 seconds at 1080p and 10 seconds at 1440p or 4K; Pro reaches 10 seconds.

Can LTX 2.5 create synchronized audio?

Yes. LTX generates audio and video jointly, using separate streams linked by bidirectional cross-modal attention. Supported pipelines produce sound with the picture and offer guidance for each modality. Write dialogue, delivery, ambience, effects and music into the same chronological scene. As with any generative model, review multi-speaker voices and long sequences before using them in production.

Is LTX 2.5 open source, and can I use it commercially for free?

The weights plus the inference and training code are public under the LTX-2.x Community License. Lightricks says organisations with under $10 million in total annual revenue can use it commercially and in production at no cost, subject to the full licence; bigger organisations need a paid commercial agreement. Passing on fine-tunes and redistribution can come with extra conditions, so base deployment decisions on the licence itself, not on a summary like this one.

What hardware do I need to run LTX 2.5 myself?

The recommended split distilled download weighs in at around 66 GiB before optional extras. The official Python setup wants Python 3.12+, CUDA 12.7+ and a current PyTorch install. FP8 casting, NVFP4 on Blackwell cards, offloading to CPU or disk, tiling and alternative attention backends can lower memory use, but high-resolution audio-video generation remains demanding GPU work, not something for a light desktop.

What does the official LTX 2.5 API cost?

LTX's launch page quotes official API prices per generated second: $0.09 at 720p, $0.15 at 1080p, $0.19 at 2K and $0.37 at 4K. Those are LTX's own API rates, not MuseGen credit prices. The MuseGen credit cost for your settings appears in the generator before you submit.

Where the LTX 2.5 details come from

Capabilities and media on this page were checked against the developer's official product pages, announcements and documentation.

LTX — LTX 2.5 model launch LTX — August 2026 release notes Lightricks/LTX-2.5 on Hugging Face Lightricks/LTX-2 on GitHub

Make your next clip with LTX 2.5

Head back to the full LTX 2.5 AI video generator at the top, add a prompt or your references, and turn the next idea you have into a finished clip.

Generate now See every model