---

name: inference-pipelines description: "Helps agents select LTX-2 inference pipelines, build safe CLI/Python invocation plans, choose checkpoint layouts, and diagnose conditioning, HDR, retake, audio, Dub-It, and DFR generation issues." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

LTX-2 inference pipelines

Use this sub-skill when the user needs to choose or prepare an LTX-2 inference pipeline for text-to-video, image-to-video, video-to-video IC-LoRA, keyframe interpolation, audio-to-video, retake, HDR/EXR, Dub-It, text-to-audio, or DFR detail-fidelity rendering. It is for building correct commands and Python calls, validating local asset layouts, and troubleshooting before a generation run.

Do not use this sub-skill for low-level tensor/model construction, package installation, accelerator tuning, dataset work, or training. Route those requests to the sibling sub-skills named in Routes.

Fast workflow

1. Identify the requested conditioning source and output type, then choose a pipeline with pipeline-selection.md.
2. Decide the checkpoint layout: LTX-2.5 split components or LTX-2.3/legacy monolith. Use cli-reference.md for required flags and recipes.
3. Validate prompt, frame count, dimensions, input media, and HDR/EXR rules with conditioning-and-hdr.md.
4. For programmatic use, map the choice to the public class and signatures in python-api.md.
5. If preparing a command without running generation, use the bundled helpers:
   • scripts/inspect_pipeline_cli.py prints or runs --help for known ltx_pipelines modules.
   • scripts/build_distilled_command.py builds a safe LTX-2.5 split DistilledPipeline command and validates local asset paths by default.
6. If a planned run fails or looks risky, diagnose with troubleshooting.md.

Public runtime safety

• The references here are self-contained distilled operating knowledge. Do not open original repository docs, examples, tests, or source scripts during normal runtime use.
• The bundled scripts are inspection/build helpers only. They do not download models, train, generate video/audio, use credentials, contact the network, or delete user data.
• Always use local model and media paths supplied by the user. Avoid commands that implicitly download checkpoints.

Routes to sibling sub-skills

• core-components: custom model builders, schedulers, denoisers, VAE internals, tensor shapes, latent/state components, or hand-authored diffusion loops.
• performance-backends: installation, CUDA/ROCm/MPS setup, optional NATTEN/FlashAttention/DiffVAE backends, quantization tradeoffs, compile tuning, multi-GPU execution, or memory benchmarking. This sub-skill only names the relevant inference flags.
• data-preparation: media/dataset preprocessing, caption manifests, training datasets, or batch asset conversion.
• training-workflows: LoRA/full training, IC-LoRA training, trainer config files, or checkpoint export.