LTX-2 Train Model — IMPERIAL Skill

Upstream

  • Repository: Lightricks/LTX-2
  • Official agent skill: .claude/skills/train-model
  • Trainer package: ltx-trainer 1.2.0
  • Python: >=3.10
  • Native runtime: Linux + CUDA; 32 GB VRAM low-memory tier, 80 GB recommended standard tier.
  • Full fine-tuning: multi-GPU/FSDP, high-end GPUs required.

IMPERIAL binding

  • Skill ID: LTX-2-Train-Model
  • Skill version: 1.2.0-imperial.1
  • Val Town/Deno state: ACTIVE_BOUNDED_ORCHESTRATION
  • Native trainer state: EXTERNAL_CUDA_GPU_WORKER_REQUIRED
  • Authority: SKILL_DOES_NOT_GRANT_AUTHORITY

Mission

Orchestrate safe LTX-2/2.3/2.5 training plans, dataset preparation, mode selection, config generation, hardware readiness checks, preprocessing plans, LoRA/full-finetune routing, validation plans and evidence tracking.

Supported training intents

T2V, T2A, I2V, V2V/extension, A2V, V2A, IC-LoRA, video/audio inpainting/outpainting and joint audio-video conditioning where supported by the current upstream trainer.

Hard gates

  1. No heavy preprocessing/training before explicit approved training plan.
  2. No mutation outside a dedicated training workspace without approval.
  3. No native-training claim without Linux+CUDA+compatible Python+model files+text encoder+GPU evidence.
  4. No quality/performance claims before measured validation.
  5. Dataset rights/provenance must be recorded before training.
  6. Current LTX-2.x Community License must be reviewed before commercial deployment/training; commercial entities at or above the upstream revenue threshold require the appropriate paid commercial license.
  7. No silent checkpoint/model downloads that incur cost or large storage usage.
  8. No autonomous paid GPU provisioning.
  9. No training on private/user data outside authorized scope.
  10. No model/output promotion to production without post-train validation evidence.

Default recommendation

Prefer LoRA for first experiments unless the objective explicitly requires full fine-tuning and the hardware/license/cost gates are satisfied.

Runtime truth boundary

SKILL_BOUND != GPU_TRAINER_RUNNING. PLAN_READY != PREPROCESSING_COMPLETE. PREPROCESSING_COMPLETE != TRAINING_COMPLETE. TRAINING_COMPLETE != QUALITY_PASS. ACTIVE_BOUNDED_ORCHESTRATION != NATIVE_LTX_TRAINER_LIVE.