Load Video + Components MXD

Core ComfyUI's video path is a two-step: Load Video hands you a VIDEO object, then Get Video Components cracks it open for the frames, audio, fps and bit depth. If what you actually wanted was the frames - to feed an image-to-video prep node, a VAE Encode, or a frame-trimming chain - the first node was a formality. This is the pair welded into one, with every component already on its own output.

What it does

You point it at a video in ComfyUI's input folder (or hit the upload button on the widget, since the file combo is an upload combo). It opens the file, decodes the components, and reads the stream's bit depth straight out of the container header. It also overrides ComfyUI's change detection with the file's modified time, so replacing a file with a same-named copy actually re-triggers the node instead of serving you a cached run.

There's a sibling, Load Video MXD, built on the same base class - that one gives you video plus video_path and keeps a frozen widget contract for older workflows. Same loader, different outputs; you don't install anything extra for it.

Inputs and outputs

One input: file. That's it. The only things to know are that it must live in ComfyUI/input (the browser lists that folder) and that only real video containers are accepted - .mp4, .mov, .mkv, .webm, .avi. Hand it a GIF or a folder of PNGs and it refuses with This node only accepts video files ('.avi', '.mkv', '.mov', '.mp4', '.webm'). which is about as friendly as a validation error gets.

Five outputs, and they're the useful part:

• video - the VIDEO object, if you want to pass the file along untouched or into Combine Videos MXD.
• images - the decoded frames as an IMAGE batch. This is the one you'll use most: images to frame tools, to a VAE Encode, to an I2V prep node, straight into the pipeline.
• audio - the AUDIO stream, ready for Create and Save Video MXD's audio input or any audio node.
• fps - a float, so you can keep timing straight when you re-mux or generate a next segment that has to match.
• bit_depth - an int read from the stream. Useful when you're round-tripping: the save path only reuses a stream as-is when the requested depth matches the source, so knowing you're holding a 10-bit file tells you whether an 8-bit export is silently downgrading it.

Install

ComfyUI Manager: search Maxed Out, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

Restart ComfyUI afterward. The pack has no pip dependencies - it leans on the PyAV that ComfyUI already ships, which is a nice change from the usual "install three nodes, break your environment" routine.

The gotcha that matters: decoded frames are huge

images isn't a reference to the file - it's the whole clip decoded into a tensor batch. A 10-second 1080p clip at 30fps is 300 frames, and float32 frames are in the region of 6MB each, so you're holding something like 1.8GB for the loader's output alone, before any model touches it. Long clips are how people discover their "small" workflow is out of memory. Trim before you load if you can, or load and immediately drop the head and tail frames with the pack's Remove Frames MXD - the same frame-count hygiene that makes Wan 2.2 and LTX segment lengths line up.

Other things that bite

The node isn't in the list at all. The video half of this pack sits behind a comfy_api availability check. If comfy_api.latest won't import on your ComfyUI build, none of these register, and the only clue is [ComfyUI-MaxedOut] comfy_api not available in wan22.video_ops in your console. Update ComfyUI. Reinstalling the pack won't help.

"File not found" when you can see the file. It resolves against ComfyUI's input directory - not output, not your downloads folder. Move the clip into ComfyUI/input or use the upload button.

It's a loader, so it does no work for you. No trimming, no resizing, no fps conversion - five clean outputs and the file's own metadata. Deciding what to do with those frames is the rest of your graph's job, which is honestly the right division of labour: plumbing with no opinions is easier to debug.