Load Video (Advanced)

A clip becomes frames, audio, and metadata in one pull

By WASasquatch·Created 3 years ago·Updated 4 days ago· 1,844

video
images
audio
metadata

◄file▾►

◄num_frames16►

◄strategyuniform►

◄nth1►

◄seed0►

◄target_fps0.00►

◄resize_modefit and pad►

◄width0►

◄height0►

◄start0►

◄end-1►

◄max_size1024►

◄interpolationlanczos►

◄alignmiddle center►

◄pad_color#000000►

◄channelsRGB►

Load Video (Advanced) is the one video-input node that hands you the whole clip in a single pull: the frames as an image batch, the video itself with its audio, and a metadata object describing the read. No decoding steps, no ffmpeg incantation, no "extract every frame to a folder first". You pick a file, you get a batch. The "(Advanced)" is there because it sits in a suite with a Load Video (Upload) sibling for remote files and a Video Info node for reading clips without the batch overhead - this one is the full-fat version.

It's part of WAS Node Suite (menu WAS Suite/IO), whose v3 rewrite builds video I/O on ComfyUI's own media stack - no packages for you to install, which is rarer in video land than it should be.

How it works

The file menu lists videos across ComfyUI's input, output, and temp folders - each entry tagged (clip.mp4 [input]) so you never guess where a file lives. There's an upload button that drops a file into input and selects it, and an in-node player shows what's selected, so you can check the clip before it ever reaches the sampler.

The frame selection is the part worth understanding, because a clip can hold thousands of frames and a batch is one tensor in memory. num_frames defaults to 16 and strategy decides which 16 you get:

uniform - evenly spaced, a contact sheet of the whole clip
head / center / tail - the start, middle, or end, which actually play as a run
random - a seeded pick
every_nth - every nth frame

nth thins the pool first (2 means "every other frame") and applies to every strategy. start/end limit the range the strategy may choose from, with negatives counting back from the end (-60 starts sixty frames before the last). Set num_frames to 0 to take every frame in range up to a 4096 ceiling - useful, memory-expensive.

target_fps is the subtle one: 0 keeps the file's own rate; a lower rate drops frames and a higher one repeats them so the clip runs for the same duration either way. Set it to 8 or 16 to match what a video model expects to see. Sizing works like the suite's image loaders - width/height at 0 keep the encoded size, and max_size (default 1024) caps the longest edge so a 4K clip doesn't try to fill VRAM in one batch.

What comes out

Four wires, and they're all useful at once:

video (VIDEO) - the kept frames with their sound, at the output rate. Wire into the suite's Save Video (Advanced).
images (IMAGE) - the same frames as one batch, in playback order. This is the wire for anything that samples per-frame, or for an image-to-video pipeline that wants a first/last frame pair.
audio (AUDIO) - the sound for the kept span. Empty when the file is silent - check has_audio in the metadata before wiring it into a save.
metadata (WAS_VIDEO_METADATA) - rate, frame count, size, duration, bit depth, whether there's sound. Video Metadata unpacks any of them to a number you can act on.

Installing it

Standard WAS Node Suite install: ComfyUI Manager → "WAS Node Suite v3", or

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

restart, done. ComfyUI 0.14.0+ and Python 3.10+ are the requirements; the pack ships no pip dependencies, and video decoding rides ComfyUI's own media stack.

Where people get burned

Two things trip people up. First, defaulting to 16 frames: if you expect the whole clip on the images wire, you get a 16-frame sample instead - raise num_frames or set 0, and mind VRAM. Second, target_fps and strategy are easy to confuse: uniform spreads across the whole clip (great for picking diverse stills), while head/tail are what you want when you need consecutive frames that actually animate. And if the audio output comes back empty on a video you know has sound, it's usually that the kept frame range doesn't include the audio track - check the metadata before assuming the file is silent.

CategoryWAS Suite/IO

Inputs (16)

Name	Type	Default	Description
file	COMBO		Which video to read. Each entry carries the folder it sits in: `clip.mp4 [input]`, `render.mp4 [output]`, `scratch.mp4 [temp]`. The button below uploads one into input and selects it, and the player shows what is selected.
num_frames	INT	160–4096	How many frames to keep, chosen by the strategy below. 16 by default, because a clip can hold thousands and a batch is one tensor in memory. 0 takes every frame in the range, up to the 4096 ceiling.
strategy	COMBO	uniform	How num_frames are chosen. uniform = evenly spaced; head = first; center = middle; tail = last; random = a seeded pick; every_nth = every nth. uniform gives a contact sheet of a whole clip, head gives a run that plays.
nth	INT	11–16384	Step between the frames the strategy may choose from. 1 uses every frame; 2 thins to every other one first, so `head` takes the opening of the clip on alternate frames. It applies to every strategy.
seed	INT	00–18446744073709550000	Seed for random, so a re-run keeps the same frames. Ignored by the other strategies. Any whole number; `0` is as good a seed as any.
target_fps	FLOAT	0.000–240	Rate the frames come out at. 0 keeps the file's own. A lower rate drops frames and a higher one repeats them, so the clip runs for the same time either way. Set it to match a model that wants 8 or 16 fps.
resize_mode	COMBO	fit and pad	How each frame meets the size below. `fit and pad` keeps the whole frame and pads the rest, `fill and crop` fills the size and trims the overhang, `stretch` distorts to fit, `crop or pad` never resamples.
width	INT	00–16384	Width every frame is brought to. 0 takes the width the file was encoded at, which is what loads a clip at its own size.
height	INT	00–16384	Height every frame is brought to. 0 takes the height the file was encoded at.
startopt	INT	0-16384–16384	First frame to consider, counting from 0 through the file's own frames. Negative counts back from the end, so -60 starts sixty frames before it.
endopt	INT	-1-16384–16384	Last frame to consider, inclusive. -1 is the final frame, which is the whole clip together with a start of 0.
max_sizeopt	INT	10240–16384	Longest edge the derived size is held to, keeping the aspect. Only read when width and height are 0, which is where a 4K clip would otherwise fill memory. 0 lifts the cap.
interpolationopt	COMBO	lanczos	Resampling filter. `lanczos` is the sharpest for a downscale.
alignopt	COMBO	middle center	Which part of a frame survives a crop, and which side carries the wider bar of a pad.
pad_coloropt	STRING	#000000	Fill for space a frame does not cover. Any Pillow colour.
channelsopt	COMBO	RGB	Channels the image batch carries. `RGBA` keeps the pad transparent. The video output is always colour, since a video carries no transparency.

Outputs (4)

Name	Type	Description
video	VIDEO	The frames that were kept, with their sound, as a video at the rate below. Wire it into Save Video, or into any node taking a VIDEO.
images	IMAGE	The same frames as one image batch, in playback order, every one at the same size.
audio	AUDIO	The sound playing under the frames that were kept, from where they start and for as long as they run. Empty when the file is silent, so read has_audio before wiring this into a save node.
metadata	WAS_VIDEO_METADATA	What this read measured: the rate, the frame count, the size, the duration, the bit depth and whether there is sound, beside the same figures for the file itself. Wire it into Video Metadata to read any of them as a number.