ITL Multi Video Loader

Video workflows in ComfyUI burn VRAM, and a big chunk of that burn is loading: decode a 30-second clip into frames and you've already spent time and memory before a single sampling step. The ITL Multi Video Loader's design philosophy is the opposite of most loaders - don't decode anything until you have to. Drop up to 8 clips on one node and each gets a video_N and an audio_N socket, with the video handed back lazily: no frames decoded at all unless you ask for it.

How it works

The force_rate field is the whole story:

• 0 (default) - every clip keeps its native frame rate, and video_N is returned without decoding a single frame. Free. This is the "load a bunch of clips, route them around, only pay when a downstream node actually touches the pixels" path.
• Any other value (up to 240) - clips are retimed to that rate by dropping or duplicating frames, while keeping their real running time (a 10s clip stays 10s). That's the same convention as VideoHelperSuite. But forcing a rate has to decode the clip's frames to do it, so it costs time and memory. Only set it when your pipeline genuinely needs every clip at a uniform rate - mixing 24, 30, and 60 fps clips into a single edit is the case that justifies it.

Audio is handled separately: each clip's audio track is decoded with av (PyAV, which ships with ComfyUI) so audio_N is available even on the cheap force_rate == 0 path. A clip with no decodable audio track yields None on that file's audio_N socket - which, like the row toggle, is the same as leaving an unplugged OPTIONAL input. Safe unless that socket feeds a REQUIRED input, in which case the node fails.

The settings and traps

• output_slots - auto (sockets appear/vanish with the file list) or pinned 1–8 to keep wires in place while swapping files. Extra pinned sockets output nothing.
• Row on/off toggle - keeps socket position, outputs None for video_N and audio_N. Equals an unplugged OPTIONAL input; don't toggle off a row feeding a REQUIRED input.
• count - enabled files actually emitted, not row count.

Reordering or removing files shifts what each socket carries; the node warns in its status line whenever a wired socket is affected.

Outputs & install

Outputs: count (INT), then per file video_N (VIDEO) + audio_N (AUDIO). The Advanced variant adds filename_N per file.

Install via ComfyUI Manager (search "ComfyUI-IntoTheLatent-Utils") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Into-The-Latent/ComfyUI-IntoTheLatent-Utils

then restart. No extra dependencies - the pack's only declared requirement is Pillow, and video decoding rides on ComfyUI's own PyAV support.

The trap to avoid: setting force_rate "just in case" when your clips are already at the same rate. You're paying for a full decode for zero benefit. Leave it at 0 until you have a real reason, and let the lazy path do its job.