🅛🅣🅧 LTXV Audio Only Model

Here's a fun one. LTX-2 is a single transformer that generates video and audio together - but it can also generate just audio, and LTXVAudioOnlyModel is the switch that puts it into that mode. Wire your LTX-2 model through it and you've got a text-to-audio (T2A) generator: type a prompt, get a sound, no video rendered and no video-model cost. Given that LTX-2's standout use case is synchronized sound and lipsync, having the audio half available on its own is genuinely handy for foley, ambience, and quick sound design.

The neat part is that it's not a separate model. It's the same weights, reconfigured. The README lays out exactly what it does under the hood: it flips off the model's run_vx, a2v_cross_attn, and v2a_cross_attn flags, so the audio is denoised with no dependence on the video latent and the video stream is skipped entirely. This mirrors the reference single-stage T2A pipeline's video=None behavior.

How it works

You pass in your loaded LTX-2 model; the node returns the same model with audio-only mode engaged. Because the model splits its input positionally into [video, audio], the sampler still technically needs a video latent sitting at index 0 - but a tiny placeholder does the job. The README's recipe: make a fixed 64×64 single-frame dummy with LTXVAudioOnlyEmptyVideoLatent (no params to tweak), join it to your audio latent with LTXVConcatAVLatent, and with this node active that dummy is never attended to and costs almost nothing. After sampling, pull the audio out with LTXVAudioVAEDecode and save it with a standard audio node like Save Audio (FLAC).

The inputs and outputs that matter

Dead simple interface:

• model (MODEL) - your LTX-2 model.

Output is the same model (MODEL), now in audio-only mode. Everything else - the dummy video latent, the concat, the audio decode - happens in the surrounding nodes the README names. To go back to normal joint audio+video generation, just remove or bypass this node.

How to install it

Per the README, use ComfyUI Manager: Ctrl+M → Install Custom Nodes → search LTXVideo → Install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

then restart ComfyUI. It's under Lightricks → audio. You need the LTX-2 model and its Gemma text encoder loaded (the model is 32GB+ VRAM, 100GB+ disk territory per Lightricks) - audio-only mode skips the video compute, but it's still the big model with the big text encoder.

Common issues & troubleshooting

"Why do I still need a video latent?" Because the model splits its input into [video, audio] positionally, so index 0 must exist. Use LTXVAudioOnlyEmptyVideoLatent for the placeholder - it's a fixed 64×64 single frame and, with this node active, never gets attended to. Don't skip it; the sampler expects something there.

No audio comes out / you get video instead. Make sure you decoded with LTXVAudioVAEDecode (which extracts the audio from the joint latent), not a video decode. And confirm this node is actually in the model line before the sampler - if it's bypassed, you're back in joint mode.

The Gemma text encoder is eating all your VRAM. This is the perennial LTX-2 complaint - the Gemma 3 12B encoder is huge (over 22GB in fp16) and caused most of the launch OOMs. Use a quantized Gemma (community fp8/4-bit repacks) or the pack's text-encoding nodes that cache/offload encodings, and consider --reserve-vram if you're tight. Audio-only doesn't make the text encoder any smaller.

Audio quality is underwhelming. LTX-2's audio improved substantially after launch - 2.3 brought a new vocoder and a filtered audio training set, and the distilled LoRA 1.1 fixed the mumbling that 1.0 produced. Be on a current checkpoint before judging the sound.