LatentToBatch

ComfyUI is happiest when every image in a batch is processed the same way, and miserable when you want to swap one item in a batch and leave the rest alone. That's the gap this node fills: it takes an existing LATENT batch, plus a single latent (or sub-batch), and replaces or inserts it at a chosen batch index. Nothing exotic, but it's the kind of glue that unblocks workflows like "generate a grid, then rerun one slot with a different seed/prompt and re-insert it."

The operation is deliberately simple. You give it the existing latent_batch, the latent to place, an index, and a mode:

• mode: replace (default) - overwrites the item(s) at index. Batch size stays the same. The index must point at an existing item.
• mode: insert - shifts existing items right and places the new latent at index. You can also use index equal to the current batch size to append. Batch grows.

The one non-obvious thing to know: this node is metadata-aware. If both inputs carry a noise_mask, they get spliced together properly (singleton masks are expanded to match the sample count), and existing batch_index metadata is preserved and spliced when present. So if you're re-inserting a latent that came out of a masked/inpaint pipeline, you don't lose the mask bookkeeping - a surprisingly common thing to break in hand-rolled batch surgery.

The constraints that bite

Because it's splicing tensors, the latent you place must match the batch's shape after the batch dimension - same spatial size, channels, dtype, and device. Trying to drop a 1024×1024 latent into a batch of 512×512 latents is a hard error, not a silent corruption. That's worth knowing up front: this is a batch surgery node, not a resize node. Resize/encode first, then splice.

Installing and using it

Part of the Skoogeer-Noise pack. ComfyUI Manager → search "Skoogeer-Noise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/Skoogeer-Noise

Restart ComfyUI. No models or keys; deps are torch, numpy, einops, pyyaml.

A natural workflow: run a KSampler with batch_size 4, take one latent out of the grid, perturb it (this pack is full of perturb nodes - mesh drag, swirl, channel transforms), then LatentToBatch it back in at the same index and decode the whole grid in one shot. The mode choice matters more than it looks: reach for replace when you want to keep grid size constant and insert when you're progressively building a batch. And if you're in a list-iteration workflow, note the pack also ships Models List for model fan-out and ListSlice/Reorder List for list surgery - the batch and list utilities are meant to be used together.