Nodes/ComfyUI/Load Video (from Folder)
ComfyUI Node Runs on cloud

Load Video (from Folder)

Every clip in a folder, in one list

By Comfy-Org·Created 4 years ago·Updated about 5 hours ago· 128,055
Load Video (from Folder)
    • videos
    folder

    Load Video (from Folder) is the batch version of the classic Load Video node. Where Load Video gives you one file and one VIDEO, this one points at a folder and hands back every supported video in it as a single list - ready to be shuffled, sampled, or fed into a training run.

    It's a deliberately dumb node, and that's the compliment. One input, one output, nothing else. If your workflow ever needs to iterate over "all the clips in this folder" - building a video dataset for ComfyUI's built-in trainer, batch-processing a folder of footage, or sampling a few frames from each clip - this is the node. It's new (part of the core video-dataset work that landed mid-2026) and still marked experimental in the source, so don't be surprised if it moves around the menu over the next few releases.

    How it works

    Under the hood it's a directory scan with a filter. You pick a folder from a dropdown - it only lists subfolders of ComfyUI/input/, the same place Load Video reads from - and the node lists the files, keeps the ones ending in .mp4, .avi, .mov, .webm, .mkv or .flv, sorts them alphabetically, and returns one video reference per file.

    The part to actually care about is lazy loading. Nothing is decoded at load time. You can dump 500 clips into your graph and ComfyUI won't touch a single pixel until a downstream node asks for frames. That's what makes a folder-scale loader practical instead of a VRAM bomb.

    Inputs and outputs

    • folder - the only input. A combo dropdown populated from the subfolders of input/. You can't type an arbitrary path, and it won't show the input root itself, so your videos need to live in at least one subfolder.
    • videos - a VIDEO list output. Wire it into anything that eats a list of videos: shuffle nodes, per-clip sampling, or the training pipeline.

    No model files involved - it's pure data loading, nothing to download.

    Gotchas

    • It only reads one level deep. Clips nested in deeper subfolders are invisible, and if nothing supported is found the node raises an error rather than returning an empty list. Flatten your folder - or use Load Video-Text (from Folder), which recurses into kohya-style repeat subfolders.
    • Order follows the filename sort. If you want clip order preserved meaningfully, prefix your filenames (001_x.mp4, 002_y.mp4, …). Alphabetical is the contract, and it's stable.
    • It's the wrong node for casual use. Previewing a single clip or prepping one image-to-video input is still Load Video's job. This one earns its keep when the count is in the dozens-plus and the destination is a dataset.
    Categoryvideo

    Inputs (1)

    NameTypeDefaultDescription
    folderCOMBOThe folder containing video files.

    Outputs (1)

    NameTypeDescription
    videosVIDEOLazy video references; frames are decoded only when needed downstream.