Video Loader

Here's the honest review up front: the Video Loader in this pack doesn't actually decode video into frames. Not yet. The source code literally contains a # TODO load videos comment sitting above the frame-loading call that was never written. What the node does do - reliably - is scan a folder, pick a video file by index or random, and hand you its path. If your workflow needs the path (to hand to a file-based tool, a ffmpeg wrapper, or a node that loads lazily), it's fine. If you expected image frames on the other end, adjust expectations.

What it actually does

It's the same machinery as the pack's Image Loader, pointed at video extensions instead: mp4, mov, avi, mkv. Given a folder_path it lists matching files, sorts by name or creation date or shuffles, optionally reverses, and selects one by video_idx - or at random with random_idx and a seed. Then it returns:

• image - always None (not implemented)
• mask - always None (not implemented)
• video_path - the STRING path of the selected file, the one output that works

That's the entire contract. The image and mask sockets exist, but they're wired to return nothing.

Inputs

• folder_path (STRING) - absolute path to the folder of videos.
• Optional: video_idx (INT, default 0), random_idx (BOOLEAN), sort_by (STRING: name / date / shuffle), reverse (BOOLEAN), seed (INT, default 42).

Install

Pack standard: ComfyUI Manager → "Sagado Nodes for ComfyUI", or

cd ComfyUI/custom_nodes
git clone https://github.com/5agado/ComfyUI-Sagado-Nodes
pip install -r ComfyUI-Sagado-Nodes/requirements.txt

Restart. No extra video libraries in the pack's requirements - which is why frames aren't loaded; decoding video needs a real dependency (PyAV, imageio-ffmpeg, or opencv) and this pack deliberately keeps its requirements to torch, pillow, numpy, and ollama.

How to actually use it today

If you're grabbing a video for img2vid or frame extraction, pair this node with a proper frame extractor downstream - feed video_path into a Load Video / VHS-style node or an ffmpeg call and you've got a working "pick a random video from this folder" pipeline. That's a legitimately useful pattern: iterate a folder of source clips, feed each path to a real loader. Just know that the image and mask sockets are decorative as of this version, and don't build a workflow that expects frames to flow straight out of it.