Nodes/ComfyUI-FBnodes/Load Latent File
ComfyUI Node

Load Latent File

Re-run from the latent instead of re-generating from scratch

By FranckyB·Created 6 months ago·Updated 3 days ago· 13
Load Latent File
    • LATENT
    file_path

    Load Latent File is the companion node to Save Video+'s "Save Latent" option: it reads a .latent file back off disk and hands you a LATENT tensor, so you can continue from an exact latent state instead of re-sampling from scratch. It's the ComfyUI equivalent of keeping a raw file in a video pipeline - you save the lossless intermediate, and if you want a different upscale or a different second pass, you reload it and go, skipping the expensive generation.

    It's part of ComfyUI-FBnodes, François Beaudry's utility pack, where the video+latent workflow is a first-class citizen: Save Video+ can write the .latent beside your video, Get Video Components+ auto-detects it, and VACE Stitcher will load .latent clips directly to skip lossy video decode. This node is the explicit, manual version of that same handoff.

    How it works

    The input is a single file_path (STRING) pointing at a .latent file. The node loads it with safetensors.torch.load_file and reconstructs a standard LATENT dict from the stored tensors. That's the whole mechanism - no VAE, no decode, no sampling. Because a latent is just the compressed representation your model samples against, "loading" it is cheap: you get the exact same starting point the video was generated from, bit-for-bit.

    Where that matters: say you generated a clip at low resolution, saved video and latent, then want to try a different refine pass or a different sampler on the same content. Without the latent you'd re-run the whole generation; with it, you load straight into the second stage. It also pairs with the LTX-specific variant, Load LTX Latent File, when your latent carries extra tensors (audio/video) that a plain loader would drop.

    The one input

    • file_path (STRING) - path to the .latent file, e.g. the file Save Video+ wrote next to your video.

    Output: LATENT, ready for a KSampler or a latent-manipulation node.

    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. The pack pulls in av and color-matcher; this node needs neither at runtime, but they install with the pack anyway.

    Gotchas

    The big one: a latent is only useful in the model's own latent space. A Wan latent won't mean anything to an LTX sampler, and even within a family, a latent from a different resolution or a different VAE isn't a valid starting point. Save the latent from the same pipeline you plan to continue in. Also, .latent files are large - they're uncompressed tensors, so budget disk space accordingly, and don't treat them as an archive format. And if you point it at a missing path, it raises a clear "Latent file not found" error rather than silently passing an empty latent - that's the behavior you want in a node whose whole job is precision.

    CategoryFBnodes

    Inputs (1)

    NameTypeDefaultDescription
    file_pathSTRINGPath to the .latent file to load.

    Outputs (1)

    NameTypeDescription
    LATENTLATENT