LatentBatchSeedBehavior

LatentBatchSeedBehavior controls how the seed is treated across a batch of latents - specifically whether each item in the batch gets its own independent noise pattern or shares one. It's a metadata node: it doesn't touch the latent's pixel values at all, it adjusts the batch_index bookkeeping that the sampler uses to derive per-item noise. Drop it between your empty latent and the KSampler and you're choosing a global behavior for the whole batch.

The two inputs: samples (the LATENT) and seed_behavior, a dropdown with two options:

• fixed (the default) - every item in the batch uses the seed pattern of the first item. The batch shares one "starting noise story."
• random - the batch_index is stripped, so each item derives its own noise from the seed independently.

Output is the same LATENT, relabeled. The sampler downstream reads the metadata this node wrote.

When fixed matters more than you'd think

"Fixed" sounds like the boring option, but it's the one doing the heavy lifting in real workflows. In video generation, the frames of a clip often live as a batch (or the batch carries the per-frame seeds), and fixed means every frame starts from the same noise relationship - which is what keeps motion coherent instead of every frame being a fresh roll of the dice. Same story with multi-view or consistency batches where you want the items to share a noise foundation. Random is what you want when the batch is deliberately a set of variations - a grid of different looks from one seed - because shared noise would make all four images come out suspiciously similar.

The gotcha

The behavior is silently dependent on whether batch_index metadata exists at all. If the latent coming in already carries a batch_index from an earlier node (say, after a batch operation), "fixed" re-stamps it to the first item's value; "random" pops it entirely, and the sampler falls back to its own default noise derivation. The practical effect is usually what you want, but it's metadata plumbing - if your results look oddly correlated or oddly varied and you didn't intend it, this is the node that did it. It's also a pure pre-sampling switch: there's nothing to preview, because it changes how noise is added, not what's there.

It ships with ComfyUI core - no install, no model files. For most single-image workflows you'll never touch it; the moment you start generating video or batch grids, it's the difference between "why are all my frames drifting" and "why are all my frames identical," and it's a two-click answer to both.