Fusion Inspector

Fusion is a black box until it isn't. Text Encode (Fusion) hands the sampler a conditioning whose visual tokens were blended from several reference images, and if the result comes out wrong you have no idea which reference bled where - was the face the wrong subject's, the background the wrong one's, or the whole mix just off? Fusion Inspector from ComfyUI-NynxzNodes exists to answer exactly that: it's a read-only view of the blend field, showing which source won which cell of the token grid, so a mushy output stops being a mystery.

It's an output node (nothing comes out of it) that you wire after the fact: take the fusion_inspect output from a Text Encode (Fusion) node and plug it in. That payload is captured inside the encode itself, so what you're inspecting is the blend that actually ran - not a reconstruction that could drift from it.

What it shows

• Hover the token grid and it tells you which source won each cell. If the face region of your result is coming from the wrong reference, you'll see it in the map before you waste a render guessing.
• Per-source shares - the overall proportion of the grid each image ended up owning, which is your ground truth on whether the blend was balanced the way you intended.
• The settings that produced it, captured at encode time - so when you come back to a workflow next week, the inspector doubles as a "what did I actually run here" record.

The optional inspector input is UI state (grid zoom, hover position, which panels are open) rather than something you wire meaningfully - you can generally ignore it.

When it's worth the extra node

The author's own framing is that the Fusion group keeps diagnostic tools as separate nodes precisely so the working nodes don't each carry a report socket - and that only pays off if you actually use them. Treat the inspector as a tuning instrument, not a permanent fixture: while you're dialing in fusion_method, visual_size or per-source strengths, keep it wired and watch the map change; once a blend is behaving, you can delete it and the encode works exactly the same.

The one genuine gotcha is temporal: the inspect payload is produced when the encode runs, so it reflects the last executed state. If you change the grid or the encode settings and then look at an inspector whose upstream hasn't re-run, you're reading stale data - let the workflow execute first, then inspect.

Install

Part of Nynxz's Custom Nodes: ComfyUI Manager → search "NynxzNodes" and install from the registry, or git clone https://github.com/Nynxz/ComfyUI-NynxzNodes into ComfyUI/custom_nodes. No pip dependencies. Cloned by hand? The whole inspector UI is frontend (web/ is gitignored), so you'd need pnpm install && pnpm build first - the registry/Manager build avoids that. Restart ComfyUI after install; update ComfyUI core if the nodes don't appear, since the pack targets the newer extension API.