LTXV Dual CFG Guider

LTX's audio-video models pack the picture and the soundtrack into one latent, and they don't want the same amount of conditioning pressure. That's the whole job of the LTXV Dual CFG Guider: it runs the sampler with one CFG scale on the video channels and a different one on the audio channels. If you've ever had a clip where the image looks great and the audio sounds like a muffled room tone, this is the knob you were missing - and the fix isn't a lie, it's a separate guidance scale.

It ships with ComfyUI core (it lives in comfy_extras/nodes_lt.py), so there's nothing to install. It's a 2026 addition, arriving with ComfyUI's native LTX-AV sampling support, and it's aimed squarely at the LTX-2 / LTX-2.3 / LTX-AV family of models that generate synchronized video and audio together.

How it works

A normal CFG guider computes uncond + (cond - uncond) * cfg for the whole tensor. This one splits the packed AV latent along its channel dimension and applies that formula twice: the video channels get video_cfg, the audio channels get audio_cfg. Under the hood it subclasses ComfyUI's CFGGuider, so when sampling it peeks at the latent to find where the audio channels begin (LTX-AV latents are nested tensors with the video and audio components stacked) and installs a custom guidance function that only kicks in on the audio slice.

Two quiet fallbacks matter. If the latent has no separate audio component - you're sampling plain video - or if you set both scales equal, it just behaves like a normal guider at video_cfg. No error, no drama. The one thing it does enforce: it forces both the conditional and unconditional passes to run even if one scale is 1.0, because the whole point is that the two halves can diverge.

The inputs that matter

Three wires and two numbers. The wires are familiar:

• model - your LTX AV model (the 22B dev or distilled, or a quant).
• positive / negative - the prompt conditioning pair, same as any guider.

The two numbers are the whole point:

• video_cfg (default 3.0) - guidance on the picture.
• audio_cfg (default 7.0) - guidance on the soundtrack.

Notice the defaults: audio starts more than twice as hot as video. That's not a typo. The audio stream in LTX AV models is the newer, weaker muscle, and it needs more pushing to actually follow the prompt - drop it toward the video value and voices get mumbly and under-conditioned, which was the single most-complained-about LTX-2 failure mode.

Wiring it in

It's a drop-in replacement for the CFGGuider slot in a sampler graph: model → this guider → SamplerCustomAdvanced → sampler → sigmas → latent. The single output, GUIDER, feeds the sampler. It also stacks with the other LTX guidance nodes - LTXV Modality Guidance and Spatio-Temporal Guidance sit in the model path, this sits in the sampling path, and they compose.

Where people get burned: they wire this up with a plain video latent and wonder why nothing changed. It only splits when the latent actually carries audio, so you need the full AV pipeline - audio VAE, merged video+audio latent - for it to do anything. And if the audio starts distorting or crackling, you've overshot audio_cfg; if it turns to mush, you've undershot. Start at the defaults, move one at a time, and listen more than you look.