LTX-2.5: Open Audio-Video Model (22B DiT)

LTX-2.5 is the current recommended checkpoint of Lightricks’ LTX-2 audio-video family. Official hosts we opened on 2026-10-13: the gated card Lightricks/LTX-2.5, the inference repository Lightricks/LTX-2, and the product site ltx.io.

This identification page hangs off open Runway alternatives. That mapping is editorial. We have no sourced quality comparison against Runway.

The GitHub README calls LTX-2 “the first DiT-based audio-video foundation model that contains all core capabilities of modern video generation in one model: synchronized audio and video, high fidelity, multiple performance modes, production-ready outputs, API access, and open access.” Hugging Face tags on the card include image-to-video, text-to-video, video-to-video, text-to-audio-video, and related audio/video pairs. Those tags are copied from the API, not ranked.

What we verified (2026-10-13)

Field	Value	Source
Hugging Face likes	7,589	HF API Lightricks/LTX-2.5
HF downloads (monthly counter)	1,724,548	HF API
Card created	2026-07-23	HF createdAt
Card last modified	2026-10-11	HF lastModified
Gated	auto	HF API
License label	other (ltx-2.x-community-license-agreement)	HF cardData
License file date	11 August 2026	LICENSE-2_x
GitHub stars / forks	9,727 / 1,544	GitHub API Lightricks/LTX-2
GitHub created	2026-01-03	GitHub API
Transformer files	ltx-2.5-22b-dev and ltx-2.5-22b-distilled	HF siblings + README
Text encoder	Gemma 4 12B (gemma4-12b-ltx-v1) bundled in the repo	GitHub README
Distilled download size	“roughly 66 GiB”	GitHub README quick start

The Hugging Face README itself is gated (HTTP 401 without a token). Facts above come from the public HF API metadata, the public GitHub README, and the public LICENSE file.

How to run (from the official README)

The repository documents uv sync --extra natten, then hf auth login and a download of the distilled transformer, Gemma 4 12B encoder, video VAE, audio VAE, and spatial upscaler. The CLI example is python -m ltx_pipelines.distilled with --num-frames 121. That is DistilledPipeline, described as the fast starting point.

DFR (Diffusion Fidelity Rendering) is the production-quality path. It uses the same distilled transformer plus a detailing IC-LoRA from a separate repository (Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler). The README says to expect longer runtime and more VRAM, not a different prompting style, and not to pass the full (dev) transformer to DFR.

Default DFR size in the README is 1024×1536 at 24 fps. UHD 4K is documented as --width 3840 --height 2176 (not 2160). forkalt has not run these commands.

The README also lists guided two-stage pipelines (TI2Vid, Keyframe, audio-to-video), IC-LoRA video-to-video, retake, HDR, and Dub-It. We do not invent which pipeline matches a given Runway SKU.

A ComfyUI Integration heading exists in the GitHub README. We have not copied unverified node names or a tested workflow here.

License

Hugging Face lists license other with the name ltx-2.x-community-license-agreement and a link to GitHub LICENSE-2_x. That file is titled “LTX-2.x Community License Agreement”, license date August 11, 2026, licensor Lightricks Ltd. It states it applies to LTX-2.5 versions released since that date.

Section 2.1 grants a limited license provided however that Entities with annual revenues of at least $10,000,000 (“Commercial Entities”) must obtain a paid license for any use other than a defined Non-Commercial Purpose, and should contact [email protected]. Section 2.2 defines Non-Commercial Purpose (personal/hobby use, or testing/evaluation in a non-production environment without payment). This paragraph is a pointer, not legal advice.

What we did not invent

• Peak VRAM and wall-clock time on a named GPU are not yet verified.
• We do not claim visual or audio quality relative to Runway, Sora, or MiniMax-H3.
• MiniMax-H3 exists on Hugging Face as a separate video model; it is not listed here because this page is an LTX-2.5 identification URL, not a video-model league table.
• Parameter count for VAEs and upscalers is not summed into the 22B transformer figure.

Prompting (official)

The README asks for a single flowing paragraph: main action first, then movements, appearance, background, camera, lighting, and sudden events, kept literal and within about 200 words. It points at an LTX blog prompting guide. Pipelines also expose an enhance_prompt flag. We have not judged whether enhanced prompts look “more Runway.”

Pipelines named in the repo

DistilledPipeline (fast; 8 predefined sigmas — 8 steps stage 1, 4 steps stage 2), DFRPipeline (production; same distilled transformer plus detailing), TI2Vid two-stage / HQ / one-stage, ICLoraPipeline, KeyframeInterpolationPipeline, A2VidPipelineTwoStage, RetakePipeline, HDRICLoraPipeline, DubItPipeline. Optimization notes in the same README: FP8 quantization (--quantization fp8-cast), optional FlashAttention 4 on Blackwell (flash-attn-4==4.0.0b9 against torch 2.9.1+cu128), and gradient estimation to cut denoising steps. Those names are copied so you can find them upstream. They are not a forkalt support matrix.

See also Video & Audio, Runway alternatives, Pika H3 Flex (post-trained MiniMax H3 adapters; local 480p/768p), Pika H3 Flex vs SenseNova 6.8 Code2Video, SenseNova 6.8 Code2Video (code-driven short video, not a pixel DiT), HyperFrames (HTML→MP4), and models.