Load Video Path (C2C)

Every video widget in ComfyUI lists files from ComfyUI/input/. That's fine for experiments and infuriating for actual work, where the plate lives on the NAS, on an external drive, or anywhere other than the one folder the widget can see. This node takes a path string instead, and it takes an image sequence or a sequence pattern too - so one node covers D:/plates/interview.mov and /mnt/nas/shot_010/####.exr.

Same output set as its sibling: IMAGE (or LATENT if you wire a VAE), frame_count, audio, video_info in VideoHelperSuite format, the lazy C2C_VIDEO handle, and mask.

What you type and why it validates

video is a plain string and the node checks it before the run: if there's nothing at that path, you get a message naming the absolute path it tried - with the quote-stripping applied, so a pasted "D:\plates\clip.mov" works. Sequences are detected by extension (EXR, HDR, PNG, JPG, TIFF, DPX) or by pattern (####, %04d, $F4, [1-10]), so a folder of frames and a shot_####.exr string are both legitimate inputs.

Everything else matches Load Video (C2C) - force_rate (0 keeps source), custom_width / custom_height (0 derives one from the other), frame_load_cap, skip_first_frames, select_every_nth, and the format preset that rounds frames onto a model's grid and snaps dimensions to its required multiple. If a clip came out shorter than you selected and you didn't ask for it, format is the first thing to check.

The extra widget here is sequence_fps. A video file carries its own frame rate; a folder of EXR frames does not. This is how you tell the loader what the sequence's rate is - 24 by default - and it matters the moment anything downstream cares about time rather than just frame order. Audio has no meaning for a sequence, so the audio output stays empty there.

The lazy decode, and why it's not just a gimmick

The node reads the container or sequence header first and returns frame_count plus a C2C_VIDEO handle without touching pixels. Actual decoding happens only if the IMAGE/LATENT output is wired to something, and then only for the frames your skip/nth/cap selection kept. The batch is also checked against free system memory before it's allocated, and if it doesn't fit you get a refusal with the numbers and a suggested frame cap rather than an out-of-memory death halfway through a queue.

That's the difference between a loader you can put at the top of a graph and one you're scared of.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Or "CustomNodePacks" in ComfyUI Manager, then restart. Decoding is via PyAV (av), which modern ComfyUI builds already include; if it's missing, pip install av in ComfyUI's own Python environment. No model downloads. Don't blanket-install the pack's requirements.txt over a working install - opencv-python, scipy, safetensors are the real hard requirements and ComfyUI already provides torch, numpy, torchvision and Pillow.

Where people get burned

Absolute paths are absolute, and ComfyUI may be running in a container, a WSL distro, or on a different machine from the one you're browsing on. /mnt/nas/... on your desktop isn't the same filesystem the server sees. When the validation says nothing's there, believe it and check from inside the server's environment.

Second: an EXR sequence is read as-is, and the pack's own pipeline treats images as normal 0–1 floats downstream. If you're loading 16-bit linear plates and expecting ACES-correct colour, you want the pack's colour-science nodes in the chain rather than assuming the loader did the transform.

Third: this is the loader the Image Mask Editor can show frame-by-frame without a run, along with the input-folder version - which is a genuinely nice pair for rotoscoping, since it means you paint masks on the real frames instead of on a stale preview.