innFactory AI Consulting from Rosenheim assesses new video-AI models for use in the DACH region, with a focus on data protection, compliance and practical suitability. This overview reflects the state as of October 2026. This assessment is based on the model card on Hugging Face, the GitHub repository with releases and licence text, the website ltx.io and the paper arXiv 2601.03233.
What is LTX-2.5?
LTX-2.5 is a video-audio generation model from Lightricks with 22 billion parameters. The weights are openly available: you can download them and run them on your own hardware. The model generates video and sound together, so speech, noises and music match the picture. Per the README, Lightricks describes LTX-2 as the “first DiT-based audio-video foundation model” that combines synchronised sound and picture, high fidelity, several performance modes, API access and open access in one model.
The release date can be verified from primary sources: the licence is dated 11 August 2026 and applies “for all LTX-2.5 versions released since August 11, 2026”. On the same day the GitHub repository published release v1.2.0 (“Support for LTX 2.5”). Per the releases page, v1.3.0 (26 August), v1.4.0 (29 September), v1.4.1 (30 September) and v1.4.2 (2 October 2026) followed.
| Attribute | Value (per model card or README) |
|---|---|
| Provider | Lightricks (licensor: Lightricks Ltd.) |
| Parameters | 22 billion |
| Tasks | text-to-video, image-to-video, video-to-video, each with synchronised sound |
| Resolution | up to 4K (3840×2176), width and height divisible by 32 |
| Length | up to 121 frames at 24 fps (about 5 seconds), num_frames % 8 == 1 |
| Text encoder | custom Gemma 4 12B |
| Variants | distilled and dev; bf16, int8 (ComfyUI), NVFP4 |
| Licence | LTX-2.x Community License Agreement |
| Requirements | CUDA 12.7 or newer, PyTorch around 2.7, Python 3.12 or newer |
The LTX-2 paper describes the architecture of the model family: an asymmetric dual-stream transformer with a 14-billion-parameter video stream and a 5-billion-parameter audio stream, connected by bidirectional cross-attention. The model card does not explicitly say whether and how LTX-2.5 changes this split, so we cite the paper only as the origin of the architecture.
Capabilities
- Synchronised sound: video and audio are generated in one model. The paper mentions dialogue, background music, foley effects and ambient sound.
- Native multishot: per the model card, connected scenes can be generated in one pass, with characters and look meant to stay consistent across cuts (“native multishot generation”).
- Resolution up to 4K: the model card names up to 3840×2176. The Hugging Face files include spatial and temporal latent upscalers (x2); the GitHub release v1.2.0 also names the
DFRPipeline, which per the release notes is meant to hold detail together better during fast motion. - Image, video and text input: besides text-to-video, image-to-video and video-to-video are possible; the Hugging Face metadata also lists audio-to-video and text-to-audio.
- Duration prediction: per the release notes, an optional duration component estimates the clip length from the description (
--auto-duration). - HDR output: per the release notes, scene-linear EXR frames can be written together with a BT.2020/HLG video, for example for colour grading or VFX.
According to media reports (VentureBeat), LTX-2.5 generates a 10-second 720p image-to-video clip on two NVIDIA GB200 in 6.8 seconds. This figure does not come from a primary source we reviewed and cannot be transferred to other hardware.
The licence in detail
LTX-2.5 is released under the LTX-2.x Community License Agreement. It is not an OSI open-source licence but a licence with a revenue threshold. The key points from the licence text:
- Below USD 10 million annual revenue: use, modification and distribution are royalty-free, subject to the use restrictions in Attachment A. The model card speaks of commercial use at no cost.
- From USD 10,000,000 annual revenue: such “Commercial Entities” need a paid Commercial Use Agreement for any use beyond narrowly defined non-commercial purposes; contact is
ltxv-licensing@lightricks.com. Affiliates are aggregated for the threshold. - Exception for testing: larger companies may use the model for testing, evaluation and non-commercial research in a non-production environment, provided no payment results from it. Revenue-generating use, use that affects end users, and training or distilling models for commercial purposes are expressly excluded.
- Derived models: fine-tuned weights, and also models trained on outputs of LTX-2.x, count as “Derivatives” and must be distributed under the same licence.
- Outputs: per the licence, Lightricks claims no rights in the generated outputs; you remain responsible for inputs, outputs and their use.
- Other provisions: the licence permits hosting as a service for third parties under certain conditions, grants the licensor update and usage-restriction rights (section 6), and is governed by the law of the State of New York.
Have the licence reviewed legally before production use, in particular the revenue threshold within a corporate group. This description is not legal advice.
Hardware and deployment
For operation, the model card names CUDA 12.7 or newer, PyTorch around 2.7 and Python 3.12 or newer. The quickstart files from the README amount to roughly 66 GiB. The weights sit in a gated repository on Hugging Face; you must accept the model terms and log in with a read token.
- ltx-pipelines: CLI and Python API, for example
python -m ltx_pipelines.distilled; the repository consists of the packagesltx-core,ltx-pipelinesandltx-trainer. - ComfyUI: integration available; the Hugging Face files include int8 checkpoints for ComfyUI.
- Diffusers: supported per the model card.
- Less memory: quantisation (FP8, NVFP4 on Blackwell GPUs) and CPU offloading; the components (transformer, text encoder, VAEs) can be loaded separately and combined.
- Trainer: per the release notes,
ltx-trainersupports LTX-2.5 training workflows.
We did not find consistent minimum VRAM figures. Plan a test on your target hardware before procuring capacity.
Data protection and GDPR
The main advantage of open weights is data sovereignty: when you run LTX-2.5 yourself, prompts, reference images and generated videos stay in your environment. This applies on your own hardware as well as with an EU cloud provider where you control the GPU instances yourself. It matters most when reference material contains personal data such as faces or voices.
Lightricks also offers a hosted API playground (console.ltx.video). EU data residency for it is not described in the documentation we reviewed; for this route we recommend clarifying the data processing agreement, processing location and subprocessors directly with Lightricks.
Regardless of the deployment route: check whether generated videos must be labelled as AI content, and observe personality and copyright rights for reference material.
Assessment
| LTX-2.5 | Veo 3.1 | Seedance 2.x | |
|---|---|---|---|
| Operation | open weights, self-hostable | hosted service | hosted service |
| Access | Hugging Face, ComfyUI, ltx-pipelines, Diffusers | Google, see Veo page | ByteDance/BytePlus, see Seedance page |
| Sound | yes, synchronised | yes | yes, synchronised |
| Licence | community licence, USD 10 million revenue threshold | proprietary | proprietary |
| EU processing | determined by you via self-hosting | see Veo page | via the BytePlus Dublin region, without an absolute routing guarantee |
OpenAI Sora (Sora 2 and the Videos API) was shut down on 24 September 2026 and is no longer an option. Other open video models have their own licences and hardware requirements; compare them case by case. We do not document a quality ranking between LTX-2.5 and the hosted models: we only have vendor claims, so we recommend a test with your own material.
Our recommendation
LTX-2.5 is interesting for DACH companies that want to generate videos with sound while keeping processing in the EU themselves: the open weights allow self-hosting without sending reference material to a third party. Against this stand your own GPU operation, a gated download and the licence with its revenue threshold. Companies from USD 10 million annual revenue should clarify a Commercial Use Agreement with Lightricks early.
Our approach: first a pilot with the distilled variant via ComfyUI or ltx-pipelines on a test GPU, in parallel the licence and data-protection review, then the decision between self-hosting and hosted services.
For an assessment of whether LTX-2.5 or a hosted service fits your data-protection, licensing and infrastructure requirements, contact innFactory AI Consulting.