LTXV Modality Guidance (A/V coupling)

Synchronized audio and video is the whole reason LTX-2 exists - talking heads, lipsync, foley, music videos. But "the model can generate sound" and "the sound lines up with the picture" are two different promises, and the second one is the one that usually needs help. LTXV Modality Guidance is the built-in dial for that second promise: it forces the audio and video streams to actually couple, so mouths move when words land.

What it does

LTX's audio-video models are cross-modal - the video branch reads the audio branch and vice versa through dedicated a2v (audio-to-video) and v2a (video-to-audio) cross-attention paths. That coupling is what puts a mouth shape on a spoken syllable. But it's a soft constraint, and left to itself the model will sometimes let the two drift apart.

This node makes the coupling a hard requirement. It clones your model and registers a post-CFG function. On every sampling step inside its active window it runs one extra forward pass with both cross-attention paths severed - the audio branch suddenly can't see the video and the video branch can't see the audio - and then pushes the final prediction away from that "modalities aren't talking" result. The math is exactly what the name says: result + (coupled_pred - severed_pred) * (scale - 1). Degrade the sync, then steer hard against the degradation.

The reference default of 3.0 comes straight from Lightricks' own reference guider. Set modality_scale to 1.0 and the node short-circuits entirely - no extra pass, no effect. That's your off switch, and it's cheaper than bypassing the node.

The inputs that matter

• model - your LTX-AV checkpoint, straight from the LTXV2 loader. It's cloned internally, so the original model stays untouched.
• modality_scale - strength of the coupling push. Default 3.0; the reference value, and a fine place to start.
• start_percent / end_percent - the slice of the denoise schedule where guidance runs. Defaults cover the whole run (0.0–1.0). For lip-sync-heavy clips you usually want the full range.

The single output is a patched MODEL, which feeds your sampler exactly like the original would. It stacks with the sibling guidance nodes - the dual-CFG guider and Spatio-Temporal Guidance - which together form ComfyUI's take on the "multimodal guider" idea: one dial for prompt adherence, one for visual stability, one for A/V sync.

Getting it

It ships with ComfyUI core - no install, no custom pack. It's brand new: it landed in the nodes_lt.py source alongside native LTX 2.5 support in mid-August 2026, so if your ComfyUI is older than that, update. It loads no model files of its own; whatever LTX-AV checkpoint you're sampling with comes in through the model input.

Where people trip

• It only does anything on the audio-video (AV) LTX checkpoints. Feed it a video-only LTXV model and there's no audio stream to couple - the node won't error, it just has nothing to work with.
• It costs a forward pass per step. One extra pass on top of your CFG passes. On LTX that's tolerable - this is the speed tier of video models - but it's real, and it stacks additively with STG's extra pass.
• Don't chase sync with a bigger hammer. If a clip drifts, raise the scale a step or two, not to 20. Over-guided output tends to get stiff and gluey - the picture stops breathing so the audio can stay locked. LTX-2's audio was already a community complaint zone (tinny, fumbles) before 2.3 fixed the vocoder; this node is the corrective, not a replacement for a good prompt.

One honest caveat: it's six days old as of writing, so the "what scale is ideal for what" lore is still being written. Start at the reference 3.0, A/B against 1.0 (off), and trust your ears more than any thread.