Load LTX Latent File

Load LTX Latent File is the LTX-aware version of Load Latent File: it reads a .latent file back off disk the same way, but it preserves the extra tensors that LTX workflows carry alongside the video latent - the separate audio branch that LTX-2 and 2.3 generate in the same pass. With a plain latent loader those extras can get dropped or flattened; this one knows the LTX payload shape and keeps it.

It's part of ComfyUI-FBnodes, François Beaudry's utility pack, which is built around the idea that a video workflow's latent is a first-class asset: Save Video+ writes .latent files next to videos, Get Video Components+ finds them automatically, VACE Stitcher consumes them losslessly, and this node is the explicit manual reloader for LTX specifically.

How it works

Input is a single file_path (STRING) pointing at a .latent file. The node loads the stored tensors and reconstructs a LATENT dict - but unlike the generic loader, it carries the LTX-style extras forward, the ones that keep the video and audio branches of the model in sync. The reason that matters: LTX's audio and video live in one latent payload, and a workflow that decodes or resamples the video branch while dropping the audio branch throws away half the generation. Reloading the full payload means a second-stage sampler or a refine pass sees exactly what the first pass produced - including the audio it's supposed to be refining against.

Mechanically it's still cheap: loading a latent is just tensor I/O, no VAE decode, no sampling. You use it the same way you'd use the generic loader - re-enter a two-stage workflow at stage two instead of regenerating stage one - just on LTX content where the payload shape matters.

The one input

• file_path (STRING) - path to the .latent file, typically the one Save Video+ wrote beside your LTX video.

Output: LATENT, ready for an LTX KSampler stage or an LTX audio/video decode.

Installing it

ComfyUI Manager - search "ComfyUI-FBnodes" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/FranckyB/ComfyUI-FBnodes.git
pip install -r ComfyUI-FBnodes/requirements.txt

Restart ComfyUI. No model downloads - it reads files you've already generated.

Gotchas

Latent spaces don't mix, and this is where it bites hardest: an LTX-2 latent is not an LTX 2.3 latent (2.3 shipped a rebuilt VAE and a new latent space), so don't try to reload a 2.x latent into a 2.3 workflow - it'll silently be garbage. Same-family, same-version, same-resolution is the rule. Also remember the file has to come from a save that included the audio branch; if you saved with an older tool that only wrote the video tensors, this node can only restore what's in the file. And as with all latent files, they're uncompressed and chunky - fine as a workflow intermediate, not an archive.