🅛🅣🅧 Stitch AV Latents

LTX-2's whole selling point is synchronized audio and video, but each generation is short, and the community's standard move for anything longer is to generate clips and join them. Splicing two MP4s in a video editor is easy. Joining two latents before the final denoise is the trickier part - and it's the part worth doing, because it lets the model invent a seamless transition across the seam instead of leaving you with a hard cut and a re-encode. That's exactly what this node is for.

What it does

It takes two AV latents (av_latent_1, av_latent_2) and stitches them, building a transition region and the masks that tell the sampler "regenerate this seam, leave the clips alone." There are two video modes:

• bridge (default) - inserts a separate transition segment between the clips, so clip → transition → clip.
• overlap_linear_video - linearly blends the tail of clip 1 with the head of clip 2 and replaces the overlap.

The important asymmetry is in the name: audio always stitches via a bridge, even in the overlap video mode. The audio bridge's length is derived from bridge_latent_frames × temporal stride / fps, converted to audio latent frames using the audio VAE's latent rate - so your sound stays aligned with the video transition instead of drifting, which is the classic failure when people extend LTX-2 clips by just concatenating.

How the transition gets filled

The bridge latents are initialized by bridge_init_mode, and this is the one input that actually changes the character of the result:

• lerp - linear interpolation between the last frame of clip 1 and the first frame of clip 2. Predictable, and the model just refines it. The safe default.
• noise - random Gaussian, the model generates the whole bridge from scratch. More freedom, more dice rolls.
• zeros - a neutral zero latent. You rarely want this one.

Masks are then built around the transition window: video_pre_frames / video_post_frames expand the mask before and after the seam in latent frames, video_slope_len softens its edges, and mask_video + video_mask_init_value toggle it on/off and set the base. Audio gets the same treatment via audio_start_time/audio_end_time/audio_slope_len - unless you flip audio_mask_bridge_only, which ignores the seconds entirely and masks exactly the inserted audio bridge, so only the bridge audio gets denoised and the audio of both clips stays untouched. That flag is the difference between "fix the seam" and "unwittingly re-roll a chunk of my finished clip."

Wire the optional vae and audio_vae in for exact sync - if you omit them it falls back on the LTX-2 defaults (temporal stride 8, ~25 audio latents/sec), which is fine until it isn't.

The inputs that matter

• stitch_mode - bridge vs. overlap; pick by whether you want a dedicated transition or a blend.
• bridge_latent_frames - transition length in latent frames (≈8 frames per video second at 24fps, given stride 8).
• bridge_init_mode - lerp for control, noise for creativity.
• audio_mask_bridge_only - protect your clip audio; enable it once the seam sounds right.

Output is a single av_latent for your sampler.

Install

Same pack, same story: ComfyUI Manager → search ComfyUI-LTXVideo-AVSplit, or

cd ComfyUI/custom_nodes
git clone https://github.com/PsypmP/ComfyUI-LTXVideo-AVSplit
pip install -r ComfyUI-LTXVideo-AVSplit/requirements.txt
# restart ComfyUI

numpy is the whole dependency list. No model files - it operates on latents you already have.

Troubleshooting

Audio desync is the community-documented villain when extending LTX-2 clips - it shows up in threads about long generations constantly. If your seam's sound drifts, the fix is usually a longer bridge (more latent frames = more runway for the audio bridge to land) and audio_mask_bridge_only on. A too-short bridge reads as a hard cut; a too-long one starts inventing content you didn't ask for. And if you're running the official ComfyUI-LTXVideo pack alongside this one, watch the near-identical node names in the search bar - make sure you're grabbing the right registration. The node also demands an LTXAVModel with NestedTensor latents, same as its sibling in this pack, so it won't do anything useful on plain video latents.