Fusion Studio

Fusion Studio is the "compose a scene" node of this pack, and it does something no other node in the family does: it's an interactive canvas where you place reference images as layers, and the same layout drives two things at once. It composites the layers into a real image and emits a fusion_input whose spatial regions mirror your arrangement - so the spatial fusion can follow the scene you built.

The canvas

The stage widget is the whole editor: drop images, position and scale each as a layer, set opacity, blend mode, fit and z-order over a canvas with its own background color or image. What you arrange is what you get in the blend - the same placement routine feeds both the composite and the fusion sources.

The three outputs

• image - the flattened composite, front-to-back with all layers blended. A real scene, ready to preview or save.
• mask - the union of all layer alphas, for when you want to feed the composite somewhere with a mask.
• fusion_input - the sources for a fusion encode node, and its shape depends on the stage's "Fuse as" toggle:
  • layers - one source per layer, each carrying its canvas box as a region, plus the background as a region-less fill. This is where the layout becomes a spatial fusion: raise Region Strength on the encode node and the layout drives the blend instead of the default geometric pattern.
  • flattened - the composite as a single source (no regions).
  • off - pass the upstream fusion_input straight through.

What makes it genuinely useful

Two details separate this from a toy. First, cutouts: a layer composites on its own alpha, so a cut-out PNG (RGBA) drops in already masked - remove your background with whatever tool you like, save transparent, and it just works. Turn a layer's cutout off to place the full rectangle even when the file has alpha. Second, there's an optional upstream fusion_input input, so a Fusion Input grid can feed the studio and its sources come first.

Where this fits: the "layers" mode is the hand-crafted alternative to Qwen3-VL Ground. Instead of asking a model where "the sky" is, you place the sky layer exactly where you want it on the canvas - which is the right call when you have a precise composition in mind and don't want to fight a grounding model's guess. If your references are already separate images with clear subjects, let Ground find them; if you're building a scene from a background plus cutout characters, Studio is the faster path.

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; the fusion encode it feeds wants a Qwen3-VL CLIP. The canvas editor is a Vue widget served from the pack's web/ build, so if the stage doesn't render, your ComfyUI frontend is probably too old to serve it.