Aligned Audio Cut Transition

The visual twin of AlignedOverlapCutTransition, doing the same seam-finding job for sound. LTX-2 and 2.3 generate audio and video in one pass, which is great right up until you try to extend a clip: now you have two audio streams that need splicing too, and if you cut them at a different frame than the video, the sound drifts off the picture. This node exists so that doesn't happen - it takes the seam index your video transition chose and applies the identical cut to the audio.

How it works. Feed it source_audio (the previous clip's audio) and new_audio (the continuation's), plus the seam_index the image transition node picked and the frame_rate both sides were generated at. The node matches channel counts (mono/stereo mismatches are handled), resamples both waveforms to a common sample rate, converts the overlap and seam from frame counts to audio samples, then concatenates: everything before the seam from the source, everything from the seam onward from the new clip. overlap_frames (default 12) tells it how much of the old tail the new clip's audio overlaps; keep it consistent with what you used on the image side.

Outputs are audio (the merged waveform) and seam_seconds - the seam expressed as a time offset, which is handy if you're doing any downstream editing that thinks in seconds rather than frames.

The wiring that matters. The whole point is alignment, so the natural setup is:

• AlignedOverlapCutTransition's seam_index → this node's seam_index
• same frame_rate on both nodes

If you set the seam to 0 and the audio still sounds off, it's usually because the two transitions disagree on frame rate or overlap - they're separate nodes with independent defaults, so be deliberate about matching them.

Install is the standard fork flow:

cd ComfyUI/custom_nodes
git clone https://github.com/Yogurt1192/LTXDirector-Extender.git

Restart ComfyUI. It uses ComfyUI's own audio utilities (comfy_extras.nodes_audio) under the hood, so there's nothing extra to install.

Where people get burned. Two edge cases the node handles gracefully, so don't panic if you hit them: a clip with no audio track at all returns the other side unchanged, and frame_rate must be greater than zero or you get a ValueError. The honest limitation to expect: the overlap-timing fixes keep audio aligned, but the README warns that extended audio can still come out speech-like with imperfect word fidelity on some prompts. Alignment fixes when things happen, not what the model decides to say.