LTX-2.5.
LTX-2.5 is an open-weight video model from Lightricks. Specs, recommended settings, hardware and licence terms, from the developer's own pages.
- Status
- Reference
- Task
- Text to video, Image to video
- Input
- Prompt, optional images or video
- Output
- Video with synchronised audio
- Developer
- Lightricks
- Parameters
- 22B
- Licence
- LTX-2.x Community License
- Released
- August 2026
What the model does.
Paraphrased from the official model card.
LTX-2.5 is the August 2026 update of Lightricks' LTX-2 audio-video model. One 22B diffusion transformer generates video and synchronised audio together from text, images or video, and it can generate several connected shots in a single pass.
It ships as one file per component: a distilled transformer with a fixed 8-step schedule and a full dev transformer for training, a fine-tuned Gemma 4 12B text encoder, a diffusion video decoder or a lighter convolutional one, an audio VAE, upscalers, and an optional duration head that sets the clip length from the prompt.
Reference page.
We don't currently serve LTX-2.5. The specs, settings and licence terms on this page come from Lightricks's own pages.
Is LTX-2.5 licensed for commercial use?
Licence: LTX-2.x Community License.
Conditional. Organisations with less than $10 million in annual revenue, counted with their affiliates, can use it commercially for free, including hosting it as a service. At $10 million or more, anything beyond non-commercial testing needs a paid licence from Lightricks. The licence also bars using it in a product that directly competes with Lightricks' own commercial products without a separate licence.
- LTX-2.x Community License, dated August 11, 2026. Free for any purpose under the Attachment A restrictions, except that entities with annual revenue of $10,000,000 or more need a paid Commercial Use Agreement (licensing@ltx.io). Revenue counts all subsidiaries and affiliates under common control. LTX-2 repo, LICENSE-2_x
- Section 2.2: such an entity can use it without the paid licence only for non-commercial purposes, such as testing and evaluation in a non-production environment. Revenue-generating use, use that affects end users, and training models for commercial use all need the paid licence. LTX-2 repo, LICENSE-2_x
- Section 3: hosting for third parties (for example as software as a service) is allowed if Attachment A is an enforceable part of your user terms, users are told about it, and recipients get a copy of the agreement. LTX-2 repo, LICENSE-2_x
- Section 3.5: a fine-tune or LoRA can't be transferred to a commercial entity until that entity holds a paid licence. LTX-2 repo, LICENSE-2_x
- Section 6: keep any watermarking, provenance metadata and disclosure features intact and comply with AI laws such as the EU AI Act; Lightricks may revoke the licence if they're removed or disabled. LTX-2 repo, LICENSE-2_x
- Attachment A, item 20: no use in any product, service or application that directly competes with Lightricks' commercial products or services, or is designed to replace them, without a separate commercial licence. Item 18: for commercial use, no training, improving or fine-tuning other models with it. LTX-2 repo, LICENSE-2_x
- Attachment A also requires disclosing machine-generated content placed in any context, and bars deepfakes without consent, military use and fully automated decisions affecting legal rights. Lightricks claims no rights in outputs. LTX-2 repo, LICENSE-2_x
This is our reading of the licence, not legal advice. Read the licence text before you rely on it.
LTX-2.5 hardware requirements.
Published by Lightricks for running the weights yourself. These are the developer's figures, not Windpaint's.
- Published VRAM
- Not stated
- Applies to
- No inference VRAM figure. The README suggests FP8 casting and CPU or disk offload for limited GPU memory.
- Neither the card nor the README gives a VRAM figure for inference. For GPU memory constraints the README suggests
--quantization fp8-cast --offload {cpu, disk}; fp8-cast downcasts the bf16 checkpoint on the fly. LTX-2 repo - Requirements on the card: Python 3.12 or newer, CUDA 12.7 or newer, and PyTorch ~2.7 recommended. LTX-2.5 model card
- The README's quick-start download (distilled transformer, text encoder, video and audio VAEs, spatial upscaler) is roughly 66 GiB. LTX-2 repo
- The
*-comfy-int8-convrottransformer and text encoder files are for ComfyUI only. The NVFP4 distilled transformer runs in ComfyUI, or in ltx-pipelines with--quantization nvfp4-prequanton Blackwell. LTX-2.5 model card - The README says the DFR pipeline takes longer and uses more VRAM than the distilled pipeline. On Hopper GPUs it suggests the FlashAttention 3 wheel and
--quantization fp8-scaled-mmwith fp8 checkpoints. LTX-2 repo - The LTX-2 Trainer (LoRA, full fine-tuning and IC-LoRA) recommends Linux with CUDA and an NVIDIA GPU with 80GB+ VRAM for its standard config. Its low-VRAM config, with INT8 quantization, targets 32GB cards such as the RTX 5090. LTX-2 repo, README.md
LTX-2.5 native resolution.
What the developer's own code and examples use.
- Width and height must be divisible by 32, and the frame count must satisfy
num_frames % 8 == 1(1, 9, 17 … 121). LTX-2.5 model card - The DFR pipeline defaults to 1024 × 1536 at 24 fps. UHD 4K is
--width 3840 --height 2176, not 2160. LTX-2 repo - The card's Diffusers example runs stage 1 at 960 × 544 with 121 frames at 24 fps, and stage 2 at twice that resolution. LTX-2.5 model card
- Leave out
--num-framesand the optional duration head predicts a clip length from the prompt. LTX-2.5 model card
Recommended settings for LTX-2.5.
Steps, guidance and defaults from the developer's card and code, for running the weights yourself.
- The distilled transformer uses a fixed 8-step schedule at CFG 1. The full (dev) transformer is the trainable one. LTX-2.5 model card
- The README calls DistilledPipeline the starting point: 8 predefined sigmas in stage 1 and 4 in stage 2. DFRPipeline, its production-quality path, uses the same distilled transformer plus a detailing IC-LoRA. The guided two-stage TI2Vid pipelines use the dev transformer with the distilled LoRA. LTX-2 repo
- Image-to-video adds
--image PATH FRAME_IDX STRENGTH; frame 0 conditions the first frame. LTX-2.5 model card - Two video decoders: the diffusion decoder (heavier, longer decode) and a convolutional decoder (lighter, no extra dependencies). LTX-2 repo
- LTX-2 pipelines support automatic prompt enhancement through an
enhance_promptparameter. LTX-2 repo
LTX-2.5 prompt guide.
From Lightricks's own prompt guide, prompt tooling and examples.
- The LTX-2 README: write detailed, chronological descriptions of actions and scenes in a single flowing paragraph, start directly with the action, keep descriptions literal and precise, and stay within 200 words. LTX-2 repo
- Its suggested structure: the main action in one sentence, then movements and gestures, appearances, background and environment, camera angles and movements, lighting and colours, and any changes or sudden events. LTX-2 repo
- The README's example prompt writes sound into the description: a quick sniff is heard, then the man speaks "with a deep male voice and a satisfied tone" and his lines are quoted. LTX-2 repo
- The card says well-structured, detailed prompts improve results, and points to Lightricks' "How to prompt LTX-2" guide for multishot prompting. LTX-2.5 model card
LTX-2.5 limitations.
As the developer states them.
- Not intended or able to provide factual information, and may amplify existing societal biases. LTX-2.5 model card
- Prompt following is heavily influenced by prompting style, and videos may not match the prompt perfectly. LTX-2.5 model card
- May generate content that is inappropriate or offensive. LTX-2.5 model card
Questions.
Can I call LTX-2.5 on Windpaint?
Not at the moment.
Can I use LTX-2.5 commercially?
Yes if your organisation, counted with its affiliates, has less than $10 million in annual revenue; hosting it as a service is allowed. At $10 million or more you need a paid licence from Lightricks for anything beyond non-commercial testing. Any size of company needs a separate licence to use it in a product that directly competes with Lightricks' own.
Is LTX-2.5 open source?
No, it's open-weight. The weights are on Hugging Face once you accept the terms, under the LTX-2.x Community License. That's Lightricks' own licence, with a $10 million revenue threshold and use restrictions, not a standard open-source one.
How much VRAM does LTX-2.5 need?
Lightricks doesn't give a VRAM figure for inference. For limited GPU memory the README suggests FP8 casting with CPU or disk offload, and the card lists ComfyUI-only INT8 files and an NVFP4 distilled transformer.
What's the difference between LTX-2.5 dev and distilled?
The dev transformer is the full model, used for training and by the guided two-stage pipelines. The distilled transformer runs a fixed 8-step schedule at CFG 1 and is what the distilled and DFR pipelines expect.
Can I train a LoRA on LTX-2.5?
Yes. The dev transformer is fully trainable with the LTX-2 Trainer, which recommends an 80GB+ GPU for its standard config and has a low-VRAM INT8 config for 32GB cards such as the RTX 5090.
Do LTX-2.3 LoRAs work on LTX-2.5?
The LTX-2.5 card says most LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes, with a small number of exceptions, so validate them before production use. The LTX-2 README says a LoRA only works with the model it was trained on.
Does LTX-2.5 work in ComfyUI?
Yes. Official LTX-2.5 workflow templates ship in ComfyUI, and the repo includes INT8 and NVFP4 files made for ComfyUI.
How do I prompt LTX-2.5?
Lightricks suggests one flowing paragraph under 200 words, in chronological order: the main action, then movement, appearance, environment, camera, lighting and any sudden changes. Sounds and dialogue can be written into the description.