Guider Parameters (Ray)

Guider Parameters (Ray) is the settings-block for LTX's multimodal guider - the thing that lets you tune prompt adherence, visual stability, and audio-video sync as separate dials instead of one blunt CFG slider. It doesn't generate anything itself. It builds a GUIDER_PARAMETERS object that a downstream guider node consumes, which is exactly the kind of node that confuses people until they see the wiring.

Why it exists: LTX-2 and 2.3 sample video and audio together, and Lightricks' own framing is "no more choosing between follows-the-prompt and doesn't-fall-apart." The multimodal guider separates concerns - CFG for prompt adherence, STG (spatio-temporal guidance) for visual stability, and a modality scale for how much each stream leans on the other. This node is where you declare those values for one modality at a time.

The inputs

• modality - VIDEO or AUDIO. Pick which stream this set of parameters describes. It matters because the guider applies per-modality settings, and the node errors if you add the same modality twice to one parameters object.
• cfg - classifier-free guidance strength for this modality (default 1).
• stg - spatio-temporal guidance strength (default 1), the detail/coherence lever.
• perturb_attn - on/off for perturbed-attention guidance (default true). Ties into the pack's LTX Perturbed Attention (Ray) ideas.
• rescale - result rescale (default 0.7), dampens overshoot from strong guidance.
• modality_scale - how strongly this modality is weighted versus the other (default 0). This is the audio-vs-video balance knob.
• skip_step - skip guidance for the first N steps (default 0). Handy for letting the very first steps run freely.
• cross_attn - on/off for cross-attention guidance on this modality (default true).

And the optional parameters input is the chaining mechanism: feed one Guider Parameters output into another's parameters and they merge - that's how you build one object covering both VIDEO and AUDIO. Two nodes, one wired to the other, one output.

The wiring

The GUIDER_PARAMETERS output doesn't go to a sampler. It goes into the pack's Multimodal Guider (Ray) node, which takes your positive/negative conditioning plus these parameters and produces the guider that the Ray sampler consumes. So the graph is: two Guider Parameters nodes (VIDEO and AUDIO, chained) → Multimodal Guider → sampler. If you connect Guider Parameters straight to a KSampler, nothing happens - it's the wrong socket.

The honest take

The defaults are sane: cfg 1, stg 1, rescale 0.7, perturb_attn on. The two knobs actually worth moving are cfg (if your video isn't following the prompt, raise it - LTX's weak axis is prompt adherence) and modality_scale (if the audio feels bolted on or the video is dragging the audio around). skip_step is underrated: a small value like 2-4 lets composition settle before guidance kicks in, which cuts down on the "guidance from the very first noisy frame" artifacts.

One trap: because it builds a config object rather than patching a model, forgetting the downstream guider node means the parameters sit there silently doing nothing - and since there are no error messages, the workflow just "looks different" than expected. Check the wiring before you blame the settings.

Install: ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt. No model downloads.