Nodes/raylight/Multimodal Guider (Ray)
ComfyUI Node

Multimodal Guider (Ray)

The Multimodal Guider for LTX-2's Video+Audio Generation

By komikndr·Created about a year ago·Updated 2 days ago· 409
Multimodal Guider (Ray)
  • ray_actors
  • positive
  • negative
  • parameters
  • guider
skip_blocks

LTX-2 is where Lightricks' lineup went multimodal - one DiT that generates video and audio together, rather than video that you bolt a soundtrack onto afterward. Sampling that thing is a different beast than plain video, and RayMultimodalGuider is the guider built for it in Raylight. If you're generating LTX-2 through the pack, this is the guider your KSampler wants.

The mechanism is STG's spatial/temporal skip idea, generalized to both modalities. The guider's internals (calc_stg_indexes) work out which self-attention blocks to perturb based on whether the run is video-only or video+audio, then skips them for a degraded pass and extrapolates away from it - same family trick as the plain STG guider, but aware of the audio stream. It also maintains per-step state (tracking the previous denoised video and audio) so the guidance stays coherent across the multimodal sequence.

The required inputs:

  • positive / negative - the CONDITIONING pair, as usual (negative can be empty conditioning).
  • parameters - a GUIDER_PARAMETERS stream, which comes from another node in the LTXV section of the pack (the node that configures the multimodal sampling parameters). This is how the guider learns the video/audio settings for the run.
  • skip_blocks - a string listing which layers to skip for the perturbation. Default is empty, meaning the guider uses the model's own skip-block list from transformer_options - the source reads it at init if you leave this blank. Set it only if you want to override which layers get perturbed.

Output is a single RAY_GUIDER, which wires into the Ray KSampler's guider input - not a ray_actors patch stream.

Where it fits

The Raylight LTX-2 pipeline: Ray Init Actor → LTX-2 model on the workers → build your conditioning → this guider → the Ray KSampler. The pack's README lists LTX-2/2.3/2.5 under supported models for USP, so the multimodal path is a supported citizen, not an experiment.

Install

Part of raylight:

cd ComfyUI/custom_nodes
git clone https://github.com/komikndr/raylight
cd raylight
pip install -r requirements.txt

Restart, or install "raylight" via ComfyUI Manager.

Gotchas

  • This is an LTX-2-and-up tool. There's no audio stream in LTXV 0.9.x/13B, so using this guider on those models is a mismatch - grab RaySTGGuider for the 0.9.x era.
  • It's fussier about its graph than the patch nodes: positive, negative, and parameters all have to be present, and the skip_blocks string is the one field beginners are tempted to type random numbers into. Leave it empty unless a workflow tells you otherwise - empty is the "trust the model" option.
CategoryRaylight/LTXV/guiders

Inputs (5)

NameTypeDefaultDescription
ray_actorsRAY_ACTORS
positiveCONDITIONING
negativeCONDITIONING
parametersGUIDER_PARAMETERS
skip_blocksSTRING

Outputs (1)

NameTypeDescription
guiderRAY_GUIDER