Text Encode Qwen Image Edit (Fusion)

This is the heart of the Nynxz Fusion pipeline, and honestly the node worth installing this pack for. Text Encode Qwen Image Edit (Fusion) takes a fusion_input - the collected references from Fusion Input or Fusion Images - plus your prompt, encodes each image independently through the Qwen3-VL text encoder, and blends their visual tokens on a shared spatial grid. One edit can then draw on many references at once, with you controlling how much each one gets a say.

If you've used Qwen-Image-Edit (the Apache-2.0 instruction editor that ate the mask-and-ControlNet workflow), you know it takes one input image plus a sentence. This node generalizes that: instead of one image, a grid of them, fused so a single prompt can compose from several. It's the silveroxides idea - encode several images and hand grid cells between their visual tokens - extended with a whole tuning layer.

The inputs that matter first

• clip - the Qwen3-VL 4B/8B text encoder (the one loaded for Qwen-Image, roughly 8GB). Wire the same CLIP every other node in this pack uses.
• prompt - your edit instruction, same style you'd give Qwen-Image-Edit.
• fusion_input - the collector output from Fusion Input / Fusion Images.
• visual_size - the shared grid's square-equivalent side. Higher = more visual tokens = finer fusion at more compute. Default 384; visual_aspect picks the grid's shape (auto takes it from the first image).
• fusion_method - how sources are tiled. The default checkerboard interleaves sources cell-by-cell in equal shares (good for style mixing; note it ignores strength). spatial-grid gives each source one contiguous cell laid out row by row. spatial-strength-random scatters cells randomly weighted by each image's strength - this is the mode where your strength sliders actually control coverage. There's also the block interleave (via block_size) and a seeded dither (via dither_ratio).

The blend knobs

blend_strength runs 0.0 = a hard per-cell mosaic (the original behavior) to 1.0 = fully feathered, with feather setting Gaussian softening of each source's territory - but past ~1.3 cells the checkerboard pattern washes out to a flat average, per the tooltip. strength_roll is the variety lever: seed-driven random re-weighting that shifts which image dominates each run. At 0 it's off; raise it to ~0.5, bump the seed, and each run rerolls the blend. pattern_jitter is subtler - it shuffles which cells each image owns rather than re-weighting, and jitter_mode decides whether jittered cells are reassigned or swapped (swap keeps every image's token count exactly).

The fancier tiers, in one breath: content_mode derives blend weights from the tokens themselves (saliency, energy, cross-attention - needs content_strength above 0); style_mode/style_strength is experimental style release for the anime→photoreal case, flattening the reference's style signature so a LoRA can set the look (start 0.3–0.5, keep away from 1.0); region_strength lets per-image regions (drawn on a Fusion Studio or grounded by Qwen3-VL Ground) override the geometric pattern.

Outputs and wiring

Two outputs. The CONDITIONING wire goes straight into your KSampler. The fusion_inspect wire feeds the pack's Fusion Inspector node - an interactive view of exactly which image won each token cell, worth a look the first time you blend three references and wonder why it looks the way it does. There's also an optional vae input if you want to inspect pixel-space.

Installing it

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

or search "Nynxz" in ComfyUI Manager, then restart. No pack-level dependencies, but you must have a Qwen3-VL text-encoder CLIP on disk - that's the real prerequisite. Everything here is explicitly experimental: schema and defaults shift between commits, so pin the version once you've got a blend you like.