VRGDG_VideoSplitter

VRGDG_VideoSplitter takes an IMAGE batch (which is what ComfyUI calls a video's frames) and splits it into fixed-size chunks. You tell it how many chunks you want and how many frames each chunk holds, and it hands you up to 50 labeled outputs - chunk_1 through chunk_50 - with unused ones returning empty batches. It's the "make this long clip processable" node.

The reason it matters has to do with how video models work. You can't feed LTX-2.3 an entire 3-minute video as one prompt - generation happens in short clips with strict frame-count constraints (frames divisible by 8, plus 1). Long-form workflows handle this by working in chunks. Look at the default frames_per_chunk: 97. That's not random - it's 12 × 8 + 1, an LTX-friendly length. The author picked the default that just works with the target model, which tells you everything about what this node is for.

The inputs that matter

• images - your video frames as one batch. Load the video (VHS is the usual loader), and here they come.
• frames_per_chunk - how many frames go in each chunk, default 97. This is your clip length; keep it aligned with what the model and sampler you're targeting can handle.
• chunk_count - how many chunks to produce, 1 to 50, default 2. A 3-minute clip at 25 fps is 4500 frames; at 97 frames per chunk that's 47 chunks, so you'd set chunk_count accordingly.

Then wire each chunk you actually need into downstream processing - upscale, enhance, or regenerate per-chunk. The unused chunk_N outputs stay empty, so a graph that only uses chunk_1 and chunk_2 isn't wasting memory on the other 48.

The gotcha: a wall of wires

Fifty outputs is a lot of cables, and it's the thing that makes people recoil. You don't use all of them - the node is built for flexibility, not because any sane workflow needs 50 chunks at once. Pick the count you need, wire those, ignore the rest. The empty batches on unused outputs are exactly so you can ignore them safely.

Setup

Shared pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

or Manager → search "vrgamedev", restart.

Where people get tripped up: mixing up chunk_count and frames_per_chunk - the two are independent, and getting both wrong yields chunks that don't line up with the original timeline. And remember it splits frames, not seconds: at 25 fps, 97 frames is about 3.9 seconds per chunk. If your math and the model's constraints disagree, the chunk length (not the count) is what you adjust first.