WorldFoundry
English中文

LTX-2.5 Distilled

LTX-2.5 Distilled from Lightricks: text-to-video and image-to-video; Runner parity pending; dedicated `ltx-2.5`.

Edit this pageGitHubMarkdown

LTX-2.5 is a 22B joint audio-video model. This inference-only integration exposes the official fast two-stage DistilledPipeline for text-to-video and image-to-video, including synchronized audio generation.

Method overview: LTX-2 encodes audio and video with causal VAEs and couples dual streams through audio-visual cross-attention.
Method overview: LTX-2 encodes audio and video with causal VAEs and couples dual streams through audio-visual cross-attention.

Code: Lightricks/LTX-2.

Build your run

Pipeline
LTX25Pipelineworldfoundry.pipelines.ltx25.pipeline_ltx25
TaskInteractive Video interactive-video
Environmentltx-2.5
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  ltx-2.5 \
  --pipeline.task-profile interactive-video \
  --json

What it is for

ltx-2.5 is the recorded text/image-to-video line. Use it when you want both prompt and still conditioning on the same catalog card.

The official distilled path encodes prompts with the bundled Gemma 4 12B text encoder, jointly denoises video and audio at half resolution, applies the LTX-2.5 latent spatial upscaler, performs the short second-stage refinement, and decodes through separate video and audio VAEs. WorldFoundry calls that source directly instead of adapting the incompatible LTX-2.3 monolithic loader. It is published by Lightricks.

Catalog tasks include text-to-video, image-to-video, and audio-video-generation. Catalog aliases include ltx2.5, ltx25, and ltx-2.5-distilled. Documented modalities: text and image in, video out.

Recorded evaluation entry points include VBench (Baseline text/image-to-video quality and motion evaluation).

LTX-2.5 audio-video generation from a prompt. Sibling cards cover 2.x and LTX-Video.

Runtime contract

Optional prompt, frames, fps, height, and width; artifacts generated_video and ltx-2.5.mp4.

CLI task profiles: interactive-video. WorldFoundry binds LTX25Pipeline.

Family status: Runner parity pending.

Accept the Hugging Face model terms, download the five required split components while preserving their repository folders, clone the recorded source revision, install its ltx-core and ltx-pipelines dependencies in the dedicated environment, and pass execute=True. The optional duration head enables prompt-based automatic duration.

Command construction and official split-component mapping are validated without importing upstream CUDA code.

Running in WorldFoundry

Launch worldfoundry-eval run ltx-2.5 --pipeline.task-profile interactive-video to sample the clip. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

Recorded weight repositories include Lightricks/LTX-2.5. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh ltx-2.5. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.