WorldFoundry
English中文

LTX-2.x

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

LTX-2.x from Lightricks: text-to-video and image-to-video; Verified; 6 variants.

Edit this pageGitHubMarkdown

LTX-2.x is Lightricks' current-generation video model family in this catalog, integrated natively for the distilled LTX-2 and LTX-2.3 image-to-video paths. The native route uses the official spatial upscaler between two denoising stages. Six variants are recorded (T2V/I2V/V2V for both 2 and 2.3), but only the two I2V routes are runnable today — the others are upstream capability profiles.

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.

The recorded paper is arXiv 2601.03233.

Project page: ltx.video. Code: Lightricks/LTX-2. Code: Lightricks/LTX-Video.

Build your run

Pipeline
LTX2T2VPipelineworldfoundry.pipelines.ltx2.pipeline_ltx2_t2v
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  ltx-2-t2v \
  --pipeline.task-profile default \
  --json

Variants

  • ltx-2-t2v
    StatusRunner parity pending

    text-to-video

    Runtime profileltx-2-t2v

    Pipeline bindingltx-2-t2v

  • ltx-2-i2v
    StatusRunner parity pending

    image-to-video

    Runtime profileltx-2-i2v

    Pipeline bindingltx-2-i2v

  • ltx-2-v2v
    StatusRunner parity pending

    video-to-video

    Runtime profileltx-2-v2v

    Pipeline bindingltx-2-v2v

  • ltx-2.3-t2v
    StatusRunner parity pending

    text-to-video

    Runtime profileltx-2.3-t2v

    Pipeline bindingltx-2.3-t2v

  • ltx-2.3-i2v
    StatusRunner parity pending

    image-to-video

    Runtime profileltx-2.3-i2v

    Pipeline bindingltx-2.3-i2v

  • ltx-2.3-v2v
    StatusRunner parity pending

    video-to-video

    Runtime profileltx-2.3-v2v

    Pipeline bindingltx-2.3-v2v

What it is for

Use ltx-2-i2v for the family's joint T2V/I2V inference path. The catalog records both task profiles on this page.

The I2V route is a two-stage denoise with the official Spatial Upscaler as the boundary, loaded automatically from spatial_upsampler_path (not a separate Studio refiner toggle). Assembly uses the shared NativeDiffusionPipeline, joint multi-stage runner, and optimization policies — no Diffusers pipeline or copied upstream pipeline.

Catalog tasks include text-to-video, image-to-video, video-to-video, and audio-video-generation. Documented modalities: text, image, and video in, video out. Also recorded: ltx-2-t2v, ltx-2-v2v, ltx-2.3-t2v.

Recorded evaluation entry points include AIGCBench (I2V-focused benchmark for the runnable routes).

LTX 2.3-class I2V. Not the 2.5 or original LTX-Video card.

Runtime contract

Artifacts ltx-2-t2v.mp4.

CLI task profiles: default, interactive-video, and control-video-to-video. WorldFoundry binds LTX23I2VPipeline.

Start on ltx-2-i2v. The family registers 6 catalog variants (cards above): ltx-2-t2v (text-to-video) — Runner parity pending; ltx-2-i2v (image-to-video) — Verified; ltx-2-v2v (video-to-video) — Runner parity pending; ltx-2.3-t2v (text-to-video) — Runner parity pending; ltx-2.3-i2v (image-to-video) — Verified among others. Family status: Verified.

Runner status is Verified: both the LTX-2 and LTX-2.3 I2V targets produced recorded nonblank Studio demo artifacts. T2V/V2V still need scoped wrappers and validation — don't select those variants expecting output. This entry is tracked separately from LTX-Video 0.9.x because upstream ships distinct HF repository ids.

Tracked separately from LTX-Video 0.9.x because upstream publishes LTX-2.x under distinct Hugging Face repository ids.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run ltx-2-t2v --pipeline.task-profile default. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

The in-tree LTX package is architecture only. Runtime assembly, scheduling, loading, offload, quantization, and the denoising loop stay in the shared diffusion_model infrastructure. The transformer is adapted from Lightricks/LTX-2 at d6053703e00195bc668cbd1d5eda9dc0b2e7b74a. Checkpoints and adapted source remain under the upstream LTX license. AlayaWorld's action/history model reuses this architecture from its own synthesis package.

Recorded weight repositories include Lightricks/LTX-2 and Lightricks/LTX-2.3. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh ltx-2-t2v. Add --download to fetch missing public checkpoints.

Recorded licenses include other; status: Verified.