Random Seed Batch

Random Seed Batch is one of the little scheduling utilities that ride along in the ComfyUI-Taylor-Attention pack, and it does exactly one thing: it turns one base seed into a deterministic list of N seeds. No RNG roulette, no hand-pasting a wall of integers into your workflow.

When do you actually want that? Any time you're running a batch and want each pass to get a fresh seed while the whole run stays reproducible. The pack itself uses this kind of batch to keep its Phase-1 distillation varied - a different seed per training sample means the model isn't memorizing a single image. Outside this pack it's the same trick for "generate eight variations of this prompt and pick the good one": feed the list to a batch or loop and every iteration gets its own seed without you touching anything.

How it works

The mechanism is refreshingly simple. The node seeds Python's random.Random with your base seed and draws count integers from the full 64-bit range. Same base seed plus same count gives you the identical list on every machine, every run - that's the "deterministic" in the description. Change the base seed and the whole list changes. Change the count and you get a different sequence entirely, because the RNG advances a different number of times.

The inputs that matter

Only two inputs exist, and both matter:

• seed - the base seed, default 0. This is the single number you write down if you want to reproduce a whole batch.
• count - how many seeds to generate, default 8, up to 4096.

The output is one INT list named seeds, which you'd wire into a KSampler's seed input or into a batch iterator.

Installing

Installation is the pack's standard story. Grab it with ComfyUI Manager (search "Taylor-Attention") or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/ComfyUI-Taylor-Attention

Restart ComfyUI, then install the pack's Python dependencies - this is not a zero-dependency pack, it drags in torch, einops, pyiqa, comet-ml and friends, and it needs a recent ComfyUI with the new v3 node API:

uv pip install -e custom_nodes/ComfyUI-Taylor-Attention

The README's one-liner still references the old folder name (ComfyUI-Approximate-Attention) - point uv at the folder you actually cloned. No model files are downloaded; everything this pack needs is trained by you into ComfyUI/models/approximate_attention/.

Common issues

A couple of honest caveats. The sequence is Python's Mersenne Twister, so don't expect it to line up with ComfyUI's built-in numpy-based seed widget - it's a different generator, and that's fine. Like everything in this pack the node is flagged experimental, and the only real way to break it is a count of zero or less (it raises rather than silently returning an empty list). It's a tiny tool, but once you're sweeping prompts or seeds by the dozen it quietly saves you a lot of copy-paste.