---

name: advanced-control description: "Route expert agents to STG/APG, Q8/VAE patching, latent normalization, decoder-noise control, attention and flow-edit tricks, PAG/FETA, RF samplers, inverse prediction, and utility nodes." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Advanced Control

Use this sub-skill for expert-only control over LTX model internals. It is the right place for advanced guider tuning, optional Q8 paths, latent/stat normalization, decoder-noise tweaks, attention-bank/override behavior, flow-edit, PAG/FETA, RF forward/reverse samplers, inverse-model sampling, and small utility nodes.

Route here

• STG/APG block skipping, sigma scheduling, preset selection, or expert guider parameters.
• Q8 patching, Q8 LoRA loading, or VAE patch support.
• Latent/stat normalization or decode-noise control.
• Attention bank / override, flow-edit, PAG, FETA, RF inverse/ODE samplers, or model-sampling prediction patching.
• Small helper nodes such as float-to-int or image-to-CPU conversions.

Route elsewhere

• Normal T2V/I2V/V2V graph assembly, sampler choice, or decode ordering -> core-generation
• Prompt, Gemma, conditioning, or generic guider setup -> prompt-conditioning
• IC-LoRA recipes, motion-track, HDR, audio, masks, or upscaler workflows -> specialized-workflows

Reference set

• Root install/backend notes -> model-and-backend-requirements
• STG/APG guide -> stg-apg-guiders
• Q8 and patch order -> q8-and-patches
• Experimental tricks and utility helpers -> experimental-tricks
• Troubleshooting -> troubleshooting
• Safe Q8 preflight -> q8_preflight.py

Rules of thumb

1. If the task is only about choosing a basic graph or prompt route, hand off to the sibling skill instead of staying here.
2. If the task only asks for ordinary cfg, rescale, or guider plumbing, check prompt-conditioning first unless the user explicitly asks for expert STG/APG behavior.
3. Treat Q8 and trick nodes as optional and experimental. Do not imply native execution proof unless the environment and a real run actually show it.
4. Prefer the smallest control surface that solves the problem; avoid patching internals when a standard guider is enough.

Evidence surface

• Live-node categories inspected: lightricks/LTXV, ltxtricks, and ltxtricks/attn.
• Source-inspected families: STG/APG, Q8 patching, VAE patching, latent/stat normalization, decoder noise, flow-edit, attention bank/override, PAG/FETA, RF samplers, inverse model prediction, and utility nodes.
• This sub-skill is route-verified and self-contained; it does not claim a native run for every advanced combination.