Meta Batch Manager 🎥🅥🅗🅢

This is the node you find at 3am when a 900-frame video keeps killing ComfyUI with an out-of-memory error and you've already tried everything else. The Meta Batch Manager lets you process a video that's simply too long to fit in RAM all at once, by chopping it into sequential chunks and feeding them through your graph one after another. It's the heavy-artillery option in Kosinkadink's VideoHelperSuite - you don't reach for it until the normal knobs (frame caps, loading to latents) aren't enough.

One thing to be clear about up front: it fixes RAM, not VRAM. If your problem is the sampler OOMing on the GPU, this isn't your node - lower your resolution or frame window instead. This is for when holding all the decoded frames in system memory is what's breaking you.

How it works

Normally a workflow runs once, top to bottom. The Meta Batch Manager changes that: it holds your inputs and outputs open between executions and automatically re-queues the whole workflow, handing the loader the next slice of frames each pass, until the input runs dry. Your Load Video node produces frames 0–15, the graph runs, then 16–31, and so on, with Video Combine stitching each chunk onto the growing output file.

The catch is honest and worth internalizing: each chunk is a hard window. Anything that smooths across frames - AnimateDiff's context, motion consistency - can't see across the boundary between one meta-batch and the next. So you get a subtle jump at each seam. That's a real limitation, not a bug, and it's why you use this only when the alternative is not finishing at all.

To wire it: connect the meta_batch output to at least one Input (a Load Video or Load Images) and at least one Video Combine. It needs both ends to know when to stop and where to write.

The one setting that matters

There's a single widget, frames_per_batch (default 16), and picking it is the whole skill:

• Loading as images (blue IMAGE noodles in your graph), each frame costs roughly 50MB of RAM. Somewhere in the 100–1000 range is safe depending on how much memory you have.
• Loading and combining in latent space (no image noodles - you connected a VAE), you can go way higher, 2,000–20,000, because latents are tiny.

Bigger batches mean fewer seams and less overhead, so push it as high as your RAM tolerates.

How to install it

ComfyUI Manager: search ComfyUI-VideoHelperSuite, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

then restart. No models to download; it's a utility pack.

Common issues & troubleshooting

Nothing happens / it errors immediately. The node must be connected to both an input node and a Video Combine. If either side is missing, it has nothing to iterate or nowhere to write.

Visible jumps in the output. Expected. Raise frames_per_batch so there are fewer boundaries, or, if the seams are unacceptable, this file is just too long to do seamlessly and you'll want to render overlapping segments and blend them in an editor.

Still OOMing. If it's the GPU, this won't help - that's VRAM. Connect a VAE and work in latent space, which is both lighter on RAM and lets you crank frames_per_batch far higher.