Fusion Images

The Fusion system in ComfyUI-NynxzNodes blends several reference images into one conditioning, and there are two ways to hand it sources. Fusion Input is the grid where you drop files and tune each one. Fusion Images is the other route: a plain collector with autogrow IMAGE sockets, so you wire Load Image (or anything else that outputs an image) straight in, no grid, no files-on-disk step.

Use it when your references already exist as nodes in the graph - an image that came off a previous sampler, a processed crop, a masked patch - rather than as files you'd hunt for in a folder. For fixed file references, the grid is usually more convenient; for procedural references, this is the natural fit.

How it works

In, you get one or more images on a set of auto-growing image sockets (just keep wiring inputs and new sockets appear - no chaining required). Out comes a single fusion_input that feeds Text Encode (Fusion) the same way the grid does.

Where it's deliberately simpler than the grid:

• One shared strength applies to every wired image. Tooltip honesty: the encode node normalizes across all sources anyway, so this number only matters when there's an upstream fusion_input carrying its own per-image strengths - it's the "these socket images matter this much relative to the grid images" knob, not an absolute.
• One shared fit (default contain) frames every wired image into the encode's visual grid. Per-image framing is a grid feature; here you get one mode for all.
• fusion_input optional input - an upstream Fusion Input or Fusion Images whose images come first, letting you stack collectors and grid in either order.

The strength-as-relative-prevalence detail trips people up: set 2.0 on this node and the wired images don't become "twice as present" in any absolute sense - they become more present relative to whatever the grid contributes. The fusion encode has its own modes (like spatial-strength-random) where per-source strength actually maps to coverage; whether your number there does anything visible depends on that mode, not on this node.

Install

It's in 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. If you clone by hand, remember the pack's UI is a built web/ that's gitignored - run pnpm install && pnpm build in the folder, or skip the pain and use the Manager/registry build. Restart ComfyUI afterwards (and update ComfyUI if nodes don't register - the pack uses the newer extension API).

Gotchas

The autogrow sockets are the node's UI, so the same frontend-not-built symptom (inputs not growing) points at the clone-without-build problem. And don't hunt for a per-image strength field that isn't there - if different references genuinely need different weights, that's the Fusion Input grid's job, and you can chain it in front of or behind this node freely.