---

name: ltx-video description: "Routes LTX-Video local inference, model/config selection, and direct pipeline/component diagnostics." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

LTX-Video

Use this root skill to route repository-backed LTX-Video tasks. It is self-contained: use the bundled references and scripts rather than depending on an LTX-Video source checkout.

Start safely

1. For local generation, install the package and inference extras in a Python 3.10+ environment:

   python -m pip install -e ".[inference]"
   

   When working only from this skill bundle, install a compatible published/local ltx-video package instead of using -e.

2. From this skill directory, run the no-download environment preflight:

   python scripts/check_ltx_video_env.py
   python scripts/check_ltx_video_env.py --deep-imports --json
   

   Add --require-package or --require-cuda only when that capability is a hard requirement. The checker never downloads checkpoints or runs generation.

3. Read cross-cutting troubleshooting before changing dependencies or attempting a heavy checkpoint run.

Route by intent

• Use local-inference for inference.py commands, InferenceConfig, text/image/video conditioning, prompt enhancement, output handling, and safe command construction.
• Use model-configs for choosing or validating 2B/13B, dev/distilled, FP8/bfloat16, base/multi-scale YAML configurations and checkpoint/upscaler fields.
• Use pipeline-components for direct LTXVideoPipeline/LTXMultiScalePipeline calls, ConditioningItem, schedulers, VAE/transformer loading, tensor-shape contracts, and no-download component diagnostics.

Common boundary cases

• A request to choose a YAML and then generate routes first to model-configs, then to local-inference.
• A failing CLI/config-path/media-output request stays in local-inference; a failure inside a direct scheduler, VAE, transformer, latent, or pipeline call routes to pipeline-components.
• FP8 configuration choice belongs to model-configs; optional Q8/FP8 kernel import failures also use the root troubleshooting reference.

Root references

• Repository provenance records the source snapshot, evidence paths, and refresh rule.
• Routing metadata provides the machine-readable scenario and entry-point map.
• Troubleshooting covers installation, imports, devices, downloads, media extras, and optional FP8 support.
• Development and tests separates safe checks from network-, checkpoint-, and hardware-heavy verification.
• scripts/check_ltx_video_env.py performs a safe environment/backend preflight.

Operating rules

1. Inspect and validate a command or config before starting generation; model, text-encoder, prompt-enhancer, and spatial-upscaler downloads can be large.
2. Do not treat CPU/component/config checks as proof of full checkpoint generation quality or GPU performance.
3. Prefer bfloat16 configurations unless compatible FP8/Q8 hardware and external kernels are explicitly available.
4. Keep dimensions and conditioning rules with the local-inference guidance, model-family semantics with model-configs, and direct tensor/component contracts with pipeline-components.
5. State when a check was skipped because CUDA, model cache, network access, optional media packages, or FP8 kernels were unavailable.

Non-goals

This skill does not cover LTX-2, ComfyUI workflow editing, LTX-Video-Trainer, generic Diffusers usage, LoRA training, hosted APIs, or external control repositories. Route those requests to their own project documentation or skill; do not infer their behavior from this LTX-Video checkout.