Fusion Input

Fusion Input is the file-on-disk side of this pack's multi-reference fusion pipeline. It's a grid: you drop images on it, give each one a strength, and it hands the whole set to the fusion encode node, which blends their visual tokens so one Qwen3-VL edit can draw on several references at once. If your references live in files, this is the collector you want - no wires, just drag, drop, and weigh.

The pitch is basically "IP-Adapter for Qwen image editing, except you control the blend." Instead of one reference image crammed into the prompt, you get N of them fused onto a shared spatial grid, with per-image weights. It's a genuinely different way to compose: reference A sets the character, reference B sets the lighting, reference C the environment, and you dial how much each one gets a say.

The grid, and what the widget actually does

The grid widget is the whole node - it's the only required input. Drop images anywhere on it, or use Add images / Browse to pull from input/, output/ or temp/. A few behaviors worth knowing, straight from the README:

• An image already in input/ is referenced, not re-uploaded (matched by name + size), so dropping the same file twice never copies it.
• Strength is relative prevalence, not an absolute gain. Doubling every image changes nothing; halving one hands its share of each token to whoever else contributes there. The % under each card is that source's live share of the result.
• Mute (the eye) drops an image from the blend without removing it.
• Order matters - the spatial patterns assign grid cells by source index, so image 1 and image 2 land in different cells. Drag thumbnails to reorder.
• Fit is per card, cycling contain (whole image, letterboxed - default), cover (center-crop to fill), stretch (distort to fill). The thumbnail mirrors the choice, so what you see is what the encoder gets.

There's an optional fusion_input input too - chain Fusion Input → Fusion Input to group sources across several grids, and the upstream grid's images come first.

Where the output goes

Output is a single fusion_input wire into Text Encode Qwen Image Edit (Fusion), then on to a KSampler. And note the README's reassurance: one image is a valid fusion - the blend becomes a passthrough, which is exactly what you want for single-reference encodes or style release.

Why choose this over the sibling Fusion Images node? Files on disk vs. a wire. Anything generated upstream (a sampler's output, a composite) needs to go through a Preview or Save node first to reach this grid - that's the one workflow difference to remember. The wire version skips that round trip.

Installing it

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

or search "Nynxz" in ComfyUI Manager, then restart. The pack has no Python dependencies and downloads nothing itself - but the Fusion encode it feeds needs a Qwen3-VL text-encoder CLIP (the 4B/8B loaded for Qwen-Image, roughly 8GB at the big end), so budget for that model download if you haven't got one.