Text Encode (Fusion)

Wire two reference images into an image-editing model and it reads them one after the other, picks what it likes from each, and hands you an average - a face that's neither subject, a style that's both-and-neither. Text Encode (Fusion) from ComfyUI-NynxzNodes attacks that at the level where the problem actually lives: the conditioning. Instead of letting the model average whole reference images, it encodes each source separately through the vision tower and then mixes their visual tokens one cell at a time, under a weight map you control - by spatial pattern or by content.

This is the engine of the whole Fusion group (the one node that does the real work - Fusion Input and Fusion Images just collect sources for it). It's built for the Qwen3-VL text encoder - load a qwen3vl_4b or qwen3vl_8b CLIP, the same tower Krea 2 and the Qwen-Image family use, which is why it's called "Fusion" rather than named after any one model. On a plain text encoder, this node has nothing to fuse.

What you actually set

Inputs: clip (the Qwen3-VL encoder), prompt, and fusion_input from a Fusion Input or Fusion Images node. Then a pile of tuning, of which a beginner needs about four knobs:

• visual_size (default 384) - the side length of the shared visual grid every source is fitted into. Bigger = more visual tokens = finer fusion detail, at more compute. 384 is a sane starting point; raise it when the blend is mushy.
• fusion_method - how sources are tiled across the token grid. The default spatial-checkerboard interleaves sources cell-by-cell in equal shares - good for mixing style, ignores strength. spatial-grid gives each source one contiguous block. spatial-strength-random scatters cells randomly but weighted by each image's strength - this is the mode where per-source strength sliders actually control coverage.
• fit (default per image) - how each source is framed into the grid; cover = center-crop to fill, contain = fit whole (letterboxed), stretch = distort.
• seed - feeds the randomized variety features (strength_roll, pattern_jitter). Leave it fixed and the encode caches; bump it to re-roll the blend.

Everything else is advanced but genuinely useful once you've got the basics: blend_strength/feather soften the per-cell mosaic into a graded blend; content_mode (saliency/energy/cross-attention) derives weights from what's in the tokens rather than geometry; style_mode with style_strength loosens the reference's grip on style so a prompt/LoRA can set the look - handy for "anime reference plus a realism LoRA"; strength_roll randomly re-weights the blend each run ("the one that actually moves the result" per the author).

Outputs: the CONDITIONING (wire to your sampler) and a fusion_inspect payload for the Fusion Inspector node - wire that in whenever you're tuning, because seeing which source won which cell turns blind knob-twiddling into debugging.

Install

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 deps, but hand-clones need the UI built (pnpm install && pnpm build; web/ is gitignored) - the registry/Manager install is the easy path. Restart, update ComfyUI if nodes don't show.

The trap

Fusion and the pack's Regions group share the word but do different jobs: fusion decides which reference wins a token, Regions decides which LoRA fires on one. Don't reach for Fusion to separate character identities - that's a per-token LoRA gate's job, not a visual-token blend's.