🅛🅣🅧 Multimodal Guider

Multimodal Guider is LTX-2's answer to the classic video-model dilemma: crank guidance up and the clip follows your prompt but falls apart over time; turn it down and it stays stable but ignores you. This guider splits those concerns - prompt adherence, temporal stability, and audio-video sync - into knobs you set independently, then produces a single guider your LTX sampler can use. It's the guider you want on any LTX-2 or LTX-2.3 job with sound.

It's part of ComfyUI-LTXVideo, Lightricks' official pack, and it arrived in the January 2026 drop aimed straight at the launch complaints. In Lightricks' own words, the goal was "no more choosing between 'follows the prompt' and 'doesn't fall apart.'" Because LTX-2 makes video and audio in one model, a guider that treats both streams was overdue.

How it works

Rather than exposing every setting on the guider itself, this node reads a GUIDER_PARAMETERS bundle - built by one or more Guider Parameters nodes, typically one for VIDEO and one for AUDIO chained together. It applies each modality's settings to the right stream, including the cross-modality scale that governs how tightly audio and video lock to each other. On top of that it takes a skip_blocks spec, which controls the Spatiotemporal Skip Guidance layer-skipping directly - the mechanism that sharpens motion by steering away from a layer-skipped version of the model.

The inputs that matter

• model, positive, negative - the usual guidance trio.
• parameters - the GUIDER_PARAMETERS bundle from Guider Parameters. This is where CFG, STG, modality_scale, and the VIDEO/AUDIO split actually live; the guider just consumes them. This socket is the heart of the node.
• skip_blocks - a multiline string listing which transformer blocks STG skips. Leave it at the workflow's default unless you know which layers you're targeting; this is the "sharpen motion" lever and also the easiest one to overdo.

The single output is a GUIDER, which plugs into the guider input of LTXV Base Sampler, Tiled Sampler, or Looping Sampler.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It's under lightricks/LTXV. It has no weights, but it belongs to an LTX-2/2.3 graph, so you'll need the LTX checkpoint and the Gemma text encoder in place (heavy first-use downloads, 100GB+ disk recommended by the pack).

Where people get burned

• No parameters wired in. This node is a consumer - without a Guider Parameters bundle it has nothing to apply. If it errors or behaves like a plain guider, check the parameters chain first.
• Only setting video. For an audio-video generation, chain both a VIDEO and an AUDIO Guider Parameters into the bundle. Configure just video and the audio stream runs on defaults, which is a frequent "why does my audio ignore my settings" cause.
• Heavy STG. The skip_blocks / STG path is powerful and easy to overcook - too much and hair and skin get that exaggerated, unnatural LTX look. Back it off if faces go rubbery.
• Expecting it to fix weak I2V or fast-motion smudge. The guider tunes trade-offs; it doesn't rewrite the model's known limits. LTX-2's early I2V could freeze on the input frame and fast motion still softens - 2.3 improved both, and the community's cleanup habit is an occasional Wan video-to-video pass.