Nucleus-Image Zero Conditioning

Sometimes you don't need a prompt to be meaningful - you just need a conditioning vector to exist so the graph runs. Nucleus-Image Zero Conditioning is that node: it emits a NUCLEUS_CONDITIONING of zeros without ever loading the 8.2GB Qwen3-VL text encoder. No inputs, no settings, one output. Click, done, instant.

How it works

The node just builds a (1, 8, 4096) tensor of zeros in bf16 - one token position, eight steps of sequence, 4096 channels, which is Qwen3-VL's hidden size. It's the smallest token shape that satisfies the transformer's conditioning input, and the attention mask is dropped entirely. That's the whole mechanism. It's a placeholder, not a real encoding, and the code's own docstring is refreshingly honest about it: quality may differ slightly from a real empty-string encode, and it tells you to use Text Encode (Dual) for best quality.

Where you'd actually use it

Two legit jobs:

1. Negative-conditioning stand-in. The Sampler requires a negative input even if you don't care about it. When you want raw unconditional generation or you're just testing a workflow, wiring Zero Conditioning into the negative slot satisfies the graph with zero cost and zero load time.
2. Unconditional runs. Feed it as the positive too, and you get pure prior-sampling - the model generating with no prompt steering. Mostly a debugging or curiosity move.

What it is not is a free replacement for a real negative prompt. If you've read the README's FAQ, you know this model is sensitive to having a negative wired up, and an empty-string real encode already does a slightly better job than zeros because it carries the template's actual tokens. The author's guidance is unambiguous: for actual generations, use Dual.

Inputs, outputs, install

No inputs at all. One output, NUCLEUS_CONDITIONING, which feeds the Sampler's positive or negative port. That's the entire surface area.

Install is the pack install - ComfyUI Manager (search "Nucleus-Image") or:

cd ComfyUI/custom_nodes
git clone https://github.com/a180265/Nucleus-Image-comfyui-beta

restart, done. It needs no model files of its own, which makes it handy as the first node you add to sanity-check that the pack itself loads cleanly before you download 25GB of weights.

The honest take

It's a two-minute node with a two-line purpose, and there's nothing wrong with that. Keep it around for fast negative slots and workflow stubs. Just don't kid yourself that zeros are prompt engineering - when output quality is the goal, spend the extra seconds on a real encode.