Fusion Input

Fusion Input is where the reference images for ComfyUI-NynxzNodes' Fusion system live. It's a card grid on the node: you drop images onto it, each card carries its own strength, framing and mute switch, and you drag cards to reorder them. One tidy bundle of "these are my references, and this is how much each one matters" that feeds a fusion encode node.

It's the UI-first way to gather references. The alternative sibling is Fusion Images, which collects plain IMAGE sockets instead - think of Fusion Input as the "pick from files with your mouse" route and Fusion Images as the "already wired into the graph" route. They chain either way round, so a real workflow often mixes both: a couple of fixed references on the grid plus a live socket.

What you can do on the grid

• Per-image strength. Unlike Fusion Images' single shared strength, every card here has its own. This is where "the outfit reference should matter twice as much as the pose reference" gets expressed - and, depending on the fusion method you pick in the encode node, it can control coverage of the blend rather than just emphasis.
• Per-image fit. Framing per card (contain/cover/stretch style choices), for when one reference is a wide shot and another is a tight crop.
• Mute. Kill a card without deleting it - handy when you're A/B-ing which reference actually helps.
• Reorder by dragging. Order matters because sources are numbered and blend in that order, so this is real control, not tidiness.
• Real thumbnails. Cards read from input/, temp/ and output/, so you see actual previews instead of "file 04.jpg". If an image you dragged in looks blank, it probably lives somewhere outside those folders.

The node takes an optional upstream fusion_input (its images come first, then this grid's), and outputs one fusion_input object. That object type is the pack's own - you won't find it outside the Fusion group - and its only real destination is Text Encode (Fusion).

When to use it

Reference-fusing matters most when the model you're prompting is a visual-conditioning architecture - the pack's Fusion encode is built for the Qwen3-VL text encoder family, the same tower Krea 2 and Qwen-Image-Edit use for image prompts. If you're on a model with a plain text encoder, this whole grid is moot. For models that do take reference images, the classic failure Fusion exists to fix is feeding two references in sequence: the model averages their vibes and you get neither. Fusion Input is the front door to that fix - it's where you set up the blend before the encode node decides how to mix the visual tokens.

Install

Part of Nynxz's Custom Nodes: ComfyUI Manager → search "NynxzNodes" and install from the registry, or

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzNodes

No pip dependencies. This node is 95% frontend, which is the important install caveat: the grid UI is built into web/, and that folder is gitignored - so a hand clone needs pnpm install && pnpm build before the grid renders, while the registry/Manager install ships it prebuilt. Use the registry route unless you enjoy building Vue.

Gotchas

If the grid renders but dropped images show no thumbnail, they're probably in a folder the node doesn't scan - stick to input/, temp/ or output/. And if you wire a grid and the encode seems to ignore some cards, check the mutes and each card's strength: a muted or 0-strength card still takes a grid slot and can read as "missing" downstream.