Deadline Seed

The annoying part of batching a ComfyUI job on a render farm is seeds. If every worker renders with the same seed, you get the same image a hundred times, which defeats the entire point of a batch. DeadlineSeed is the fix: it's a drop-in seed node that turns one base seed into a deterministic, per-variation sequence once the job reaches Deadline. Base seed 1000 becomes 1000, 1001, 1002, … across your variations, so each one is a genuinely different image - and reproducibly so, because the mapping is pure arithmetic.

It ships in the same pack as Submit to Deadline (doubletwisted/ComfyUI-Deadline-Plugin), and honestly it's the piece that makes the farm batch useful. The submit node gets your 50 variations onto the farm; this node makes sure those 50 variations aren't 50 copies of the same frame.

How it works

DeadlineSeed is a seed node with a secret identity. In a normal local run it does nothing but pass the seed through unchanged - feed it a seed, get that seed back. That's deliberate: it means you can leave it in a workflow that also runs on your workstation, and nothing changes.

On the farm, the Deadline worker plugin injects hidden values at render time - a task ID, a batch-mode flag, and in distributed setups a worker ID - and the node switches behavior:

• batch mode (the normal Submit to Deadline path): output = seed + task_id. With chunk_size = 1, each variation is a task, so you get exactly the README's 1000 → 1001 → 1002 sequence.
• distributed mode (multiple workers splitting one job): output = seed + worker_index + 1, so different machines diverge without colliding.

You never set any of this in the UI - the hidden inputs aren't user-facing, and the worker plugin decides the mode from the job it's rendering. The visible surface is refreshingly small.

The one input and the one output

The only input is seed, an INT with a huge default (1125899906842, the same kind of big number ComfyUI loves). You can hardcode it, or wire it from whatever seed generator you already use - this node is not opinionated about where the base seed comes from. The output is seed, also an INT, and it goes straight into the seed input of your KSampler. That's the whole wiring: seed source → DeadlineSeed → KSampler.

Installing it

Same pack, same install as its big sibling. Clone it or find "ComfyUI Deadline Submission" in ComfyUI Manager, restart, then deploy the Deadline-side plugin with the pack's PowerShell script and point workers at your portable ComfyUI installs. No extra Python dependencies - this node is pure stdlib. Full walkthrough lives in the Submit to Deadline article if you're setting up from scratch.

Things worth knowing

First: the seed only varies if the worker actually knows it's in batch mode, which means the Submit to Deadline node needs to be in the same submitted workflow. Second: these are variations, not animation frames - same vocabulary as the submit node, same gotcha. Third, the reproducibility cuts both ways: because the mapping is deterministic, re-running a job with the same base seed and same batch structure gives you the same images. That's a feature for render farms, not a bug - you can re-render variation 37 exactly. And if you see a node called DeadlineDistributedSeed in an old saved workflow, that's just the deprecated alias for this one; use the current DeadlineSeed in anything you build today.