🅛🅣🅧 LTXV Audio Only Empty Video Latent

This is one of those nodes that looks pointless until you understand LTX-2's architecture, and then it's obviously necessary. LTX-2 is a single joint model that generates video and audio together - one transformer, two streams. That's the headline feature. But it also means the model always expects a video latent at its input, even when you only want sound out of it. LTXVAudioOnlyEmptyVideoLatent is the tiny placeholder that satisfies that requirement so you can do pure text-to-audio.

If you didn't already know LTX-2 could generate audio on its own - it can. The model can run in an audio-only mode where the video stream is switched off entirely, which is how you get text-to-audio (say, generating a sound effect or a short piece of speech from a prompt) with no wasted video generation. The catch is purely structural: the model splits its input positionally into [video, audio], so index 0 has to hold something shaped like a video latent, or the whole thing errors on shape. This node provides that something.

How it works

It emits a fixed, minimal video latent - a single 64x64 frame, per the pack's own description - with no parameters at all. You join it with your actual audio latent using LTXVConcatAVLatent, then sample. When LTXVAudioOnlyModel has put the model into audio-only mode, the video slot is never actually attended to, so this placeholder adds negligible cost and never shows up in output. It's a spacer, not content. The reason the dimensions are hardcoded and hidden is exactly to stop people from fiddling with a value that doesn't matter and confusing themselves.

The inputs and outputs that matter

There's nothing to set. That's the point.

• No inputs. Drop the node in, that's it.
• latent (LATENT) out - the placeholder, which goes into LTXVConcatAVLatent alongside your audio latent.

If you were hoping for a resolution or frame-count control, there isn't one, and you don't want one - this latent is deliberately inert.

How to install it

ComfyUI Manager: Ctrl+M, Install Custom Nodes, search LTXVideo, install, restart. Or:

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

then restart ComfyUI. It's in Lightricks' official pack, under Lightricks/audio. Text-to-audio still runs the full LTX-2 model and its Gemma 3 text encoder, so the pack's hardware expectations (32GB+ VRAM, 100GB+ disk for models, models pulled on first use) apply even though the output is just sound. The pack ships a single-stage T2A example workflow if you want a wiring reference.

Common issues & troubleshooting

You're using it for image-to-video or normal generation. Don't - it's audio-only plumbing. For regular video you use a real empty latent sized to your resolution and frame count, not this fixed 64x64 stub.

Your audio came out coupled to a video, or the run tried to make video. You skipped LTXVAudioOnlyModel. This placeholder only stays free and ignored when the model is actually in audio-only mode (which switches off the video and cross-attention paths). Without that node, the model may still try to attend to the placeholder, which is not what you want.

You don't hear anything at the end. The audio doesn't come out of the video decode. Extract it with LTXVAudioVAEDecode from the joint latent, then save it with a standard audio save node like Save Audio (FLAC). If you only wired up a video decode, there's nothing there.