Batch Noise Simulate

You ran a batch of six, loved the fourth one, and now you want just that image back - bigger, or with a slightly different prompt, but with the same underlying noise so the structure carries over. If you generated that batch the normal way (one RandomNoise seed), you can't just look up what the fourth entry's seed was. Batch Noise Simulate exists to fake it: it replays the batch's noise generation and hands you the entry you ask for.

How it works

You give it the same noise source you used for the original batch - same seed, so same RandomNoise node - plus a batch_entry number. Internally it expands your single-image latent out to batch_entry copies, generates the batch noise exactly as the sampler would have, and returns just the noise for the entry you picked. From the sampler's point of view you're running a one-image batch with noise that is bit-for-bit the fourth slice of the original six.

The one constraint is in the code itself: it insists the latent you're sampling is batch size 1. Feed it a multi-image latent and it'll throw an assertion whose message reads, verbatim: "Batch Noise Simulate works with a batch size of 1, and simulates the nth entry from a larger batch." It simulates a batch; it doesn't generate one.

The inputs

• noise - the original noise source, with the original seed, from the original batch. This has to match, or the whole simulation is meaningless.
• batch_entry - which image in the batch you want (1 to 64, default 1).

Output is a single NOISE, wired into the noise socket of SamplerCustomAdvanced like any other noise source.

Using it

Set up your workflow with a RandomNoise node at the seed you used for the batch, run it through Batch Noise Simulate with batch_entry set to the winner's position, keep your latent at batch 1, and sample. That's it. Now you can re-render that entry with a different sampler, more steps, a LoRA, whatever - and because the noise is identical, you're iterating on the same image rather than rolling a fresh one.

The pairing with the pack's other batch node is worth knowing: if you'd generated the batch with Seperable Batch Noise instead, you wouldn't need this at all - every entry's seed is computable by hand, so you'd just set a RandomNoise to that seed. Batch Noise Simulate is the tool for the person who already generated the batch the normal way and can't go back. Different jobs, same family.

Installing

No models, no pip packages, nothing heavy - the pack only depends on torch and ComfyUI's own comfy_extras, both of which ship with ComfyUI. Install via ComfyUI Manager (search "cg-noisetools", pack title "Noise variation and batch noise tools"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/chrisgoringe/cg-noisetools

then restart. It's from chrisgoringe, a prolific ComfyUI node author, and like the rest of the pack it's flagged experimental - but this is one of the more mechanical nodes here, and the assert-on-batch-size-1 behavior is the only real edge it has.