Fusion Inspector

Fusion blends several reference images into one Qwen3-VL conditioning by handing grid cells to different sources. If you've ever run that and thought "why is this part of the image coming out looking like reference B?", Fusion Inspector is the node that answers you. It's the debug view: an interactive panel that shows the actual blend field your encode produced - which image wins each cell of the token grid, the per-source shares, and the settings that created it.

What it is

It's not a debugging afterthought bolted on; it's a display node, like Preview Image. Wire the encode node's fusion_inspect output into the fusion_inspect input and it renders the payload in its Vue widget every run. The key design detail is that the payload was captured inside the encode - the exact field it blended with - so what you see is literally what ran, not a re-derivation that could drift from reality.

The inputs are minimal: fusion_inspect (required) and an inspector widget that's display-only - the backend ignores its value; it exists so the frontend has a place to hold state, and it's optional so a null value never reads as a missing input. There are no outputs; this is an output node (marked as such in the schema), and its output channel is the UI itself. You drop it at the end of a wire like a Preview Image, and hover the token grid to read the blend.

When to reach for it

Whenever strength vs. coverage starts confusing you. The default checkerboard fusion ignores strength entirely (equal shares, cell by cell), while spatial-strength-random makes strength drive coverage - and it's genuinely hard to keep straight which mode is doing what across a few references. One run through the Inspector makes it visible: you'll see the 2x-strength image owning ~2x the frame, or the muted source sitting out entirely. It also shows the settings that produced the blend, which is gold when you saved a workflow weeks ago and can't remember how you got that look.

Honestly, it's the one node in the Fusion family you can skip until something looks wrong - it does no compute of its own. But when something does look wrong, it turns "why did my fusion ignore image 3?" from a guessing game into a point-and-look.

Installing it

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

or search "Nynxz" in ComfyUI Manager, then restart. No dependencies beyond the Fusion encode it inspects, which itself needs a Qwen3-VL CLIP. It's flagged experimental, and its whole job is visualization, so treat any widget oddity as the price of admission.