Load Video [BETA]

Video work in ComfyUI has a classic annoyance: getting the video into the graph. The stock Load Image won't touch it, most video loaders want files pre-split into frames, and nobody wants to babysit ffmpeg before every test. LoadVideoNode (display name Load Video [BETA]) skips all that - drag a file onto it, or hit the upload button, and it decodes the video server-side into a batch of frames with audio, fps, and frame count riding along. It's from DemonAlone-JS_Addon-ComfyUI, the JS companion pack to the author's main DemonAlone-nodes repo.

How it works

Two pieces talk to each other. The backend adds three HTTP endpoints: /uploadvideo for drag-and-drop, /get_video_metadata for instant fps/frame-count reads, and /inputvideo for the built-in preview player that sits right on the node. The decoding itself is PyAV (import av) - not ffmpeg calls you run by hand, a proper Python binding - pulling frames through frame.to_image(), normalizing them to float tensors, and stacking them into an IMAGE batch. Audio, when present, is decoded into a waveform tensor plus its sample rate.

The preview player is the underrated part: it's a real <video> element on the node, so you can scrub the source before you even run anything.

The inputs

• video - a dropdown of video files in ComfyUI's input folder (root input, same place your images go). MP4, AVI, MOV, MKV, and WebM are supported. The upload button and drag-and-drop both drop files there.
• max_frames - caps how many frames get decoded. 0 means all of them. This is your memory lever.

The outputs

Four, which is generous for a loader:

• frames - the decoded video as an IMAGE tensor, shape [N, H, W, C].
• audio - an AUDIO dict with waveform and sample_rate, or None if the file has no audio track. Wire it into VideoMakerNode or anything that consumes audio.
• fps - the video's frames-per-second as a float, so downstream nodes can time things correctly.
• frame_count - the actual number of frames loaded after max_frames cropping, not the source's total. If you crop, this reflects reality.

Installing it - and the PyAV gotcha

cd ComfyUI/custom_nodes
git clone https://github.com/DemonAlone/DemonAlone-JS_Addon-ComfyUI

Restart, or grab "DemonAlone-JS_Addon-ComfyUI" from ComfyUI Manager. Now the catch: this node imports av (PyAV), but the pack ships no requirements.txt. A fresh install that has everything else working can still hit ModuleNotFoundError: No module named 'av' the moment it loads. The fix is one line:

pip install av

(Use the same Python environment as your ComfyUI install.) The other video nodes in this pack don't need it - they encode in the browser - so this is the only one with a real external dependency hiding in it.

Where people get burned

The memory one is the big one. Decoding happens eagerly: every frame is a float tensor in RAM, so a long 1080p clip can eat several gigabytes before you've done anything with it. If a video run OOMs, max_frames is your first dial - you usually don't need all 3,000 frames to test a pipeline. The loader will happily take a 60fps file too, so check the fps output if your frame-based math looks off.

Also note audio isn't resampled or remuxed - you get the source's native sample rate, so a 48kHz file hands you 48kHz and downstream audio nodes need to cope. And while it decodes most common containers, exotic codecs will fail loudly with a "No video stream found" style error. For the everyday MP4/MOV/WebM pipeline though, it's the smoothest way in - one drag, and you're sampling frames or feeding a video-to-video model without touching a terminal.