🅛🅣🅧 LTXV Add Latents

Despite the name, this doesn't add latents together like math - it joins them end to end. LTXVAddLatents concatenates two video latents along the frames dimension, so two clips become one longer sequence. It's the straightforward, hard-cut way to stitch, and it's the building block underneath a lot of long-form LTX workflows.

The "Add" is really "append." Clip one's frames, then clip two's frames, in one latent. If you want the join to be invisible you'd reach for the overlap/crossfade transition node instead - but plenty of the time you want a clean cut, or you're assembling pieces that already flow together, and then this is the simplest tool that does the job. It also lives happily next to LTXVSelectLatents: trim each piece to length, then concatenate.

How it works

It stacks the two latents on the temporal axis and hands back the combined result. The one hard rule is dimensional: the two clips have to match on every dimension except the frames dimension - same resolution, same channel layout. The frame counts can differ (that's the whole point), everything else must agree.

The inputs and outputs that matter

There are only two inputs, and no knobs:

• latents1 (LATENT) - the first clip. Its frames come first.
• latents2 (LATENT) - the second clip, appended after.

Output is a single LATENT holding both clips back to back. Send it to your VAE decode, or into another node to keep building.

That's the entire interface. It's a plumbing node, deliberately dumb, and that's a feature.

How to install it

Through ComfyUI Manager, per the README: Ctrl+M → Install Custom Nodes → search LTXVideo → Install → restart. Manual install:

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

then restart ComfyUI. You'll find it under latent/video. It has no dependencies of its own; the LTX-2 model it serves is the thing that wants 32GB+ VRAM and 100GB+ disk.

Common issues & troubleshooting

Dimension mismatch error. This is the big one. The two latents must be identical on every axis but frames. If clip A is 768×512 and clip B is 1216×704, or they were generated at different settings, the concat fails. Generate your segments at the same resolution, or resize before joining.

A visible jump at the join. Expected - this is a hard cut, not a blend. If you want the seam smoothed, use LTXVLinearOverlapLatentTransition (or LinearOverlapLatentTransition) instead, which crossfades the boundary. Use LTXVAddLatents when a clean cut is fine or the clips already match up.

Order came out wrong. latents1 is first, latents2 is second. If your clips are reversed, swap the inputs - there's no order flag.

Length isn't what you expected. Remember these are latent frames under temporal compression, so the decoded length is a multiple of the latent frame count. Decode a quick preview to confirm the combined clip runs as long as you planned.