Fusion Images

Fusion Images is the wire-side collector for this pack's multi-reference fusion pipeline - the answer to the one limitation of its sibling, Fusion Input. That node's grid only reaches files on disk. This one takes any IMAGE socket, so a sampler output, a mask composite, or anything generated upstream can be a fusion reference without a round trip through temp/ or output/.

Think of it as an autogrowing bucket: Load Image (or any IMAGE source) → Fusion Images → Text Encode Qwen Image Edit (Fusion) → KSampler. Wire an image in and a row appears; fill more sockets and it grows, up to 16. Each row carries the same controls a Fusion Input grid card carries: a strength (the same relative prevalence - it's a share of the blend, not an absolute gain), a fit (contain / cover / stretch), and a mute, plus a live % showing that source's share of the result.

The three inputs that matter

• images - the autogrowing socket stack. Rows map to sockets positionally: the first row drives the first connected input. Wiring a new image in adds a row at full strength; unwiring one drops its row.
• strength - applied to every wired image at once. The README is upfront that the encode node normalizes across all sources, so this only really bites when there's an upstream fusion_input with its own per-image strengths.
• fit - how each image is framed into the shared grid: contain (whole image, letterboxed - default), cover (center-crop to fill), stretch (distort). The encode node's fit override can still force one mode for everything.

Behaviors worth knowing

The details that matter for sanity: muting drops a source from the blend entirely, so the spatial pattern reflows across the survivors rather than leaving empty cells. Socket order is source order, which decides which grid cells each image gets. And a batched IMAGE adds one source per frame, all sharing that socket's strength and fit - handy for image-to-video style sequencing.

There's an optional fusion_input input that gets prepended - a Fusion Input grid can feed this node and vice versa, so mix file-based and wire-based references in one blend. Output is a single fusion_input wire into the fusion encode node.

One honesty note: the README flags that per-source strength lives in the widget rather than on the sockets because a ComfyUI autogrow template only takes one input per repeat - there's no way to pair each image_N with its own strength field. So if you need per-image weights, the grid (Fusion Input) is the tool; this node is for when everything shares one strength and you just want the wires.

Installing it

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

or search "Nynxz" in ComfyUI Manager, then restart. Zero pack-level dependencies, but the fusion encode it feeds needs a Qwen3-VL CLIP (the 4B/8B text encoder from the Qwen-Image ecosystem) - that model download is the real cost of admission here.