Layout Inputs

Layout Inputs is a two-socket bundle for the Layout editor, and both sockets are optional. It exists because drawing boxes on a blank canvas is guesswork: you want the reference in front of you, and - increasingly - you want a vision model to place the boxes for you.

The fields

image (optional) is shown in the Layout editor after a run, so you can trace its composition or remix it. Wire a Load Image here instead of squinting at a separate preview panel. It's also the thing the boxes get drawn over when you're matching a real photo's layout rather than inventing one.

clip (optional) takes a text encoder that can both generate and see - a Qwen3-VL-8B loaded through CLIPLoader, as the tooltip puts it. That's the editor's agent model: ask it, and it draws the boxes for you on the canvas. Qwen3-VL is the right family here because it's the same encoder the 2026 image models use, so you likely already have it on disk for Krea 2, Qwen-Image or Ideogram setups.

One output: inputs, into the Layout node's optional socket.

Why the model doesn't eat your VRAM

The agent's CLIP is only loaded when you actually ask the agent to work. Wire Layout Inputs into Layout, run the graph, and nothing gets pulled into memory for a model you might not use. That's not an accident of implementation, it's the point of the node: the Layout editor is meant to be usable on a card that can just about hold the model you're generating with.

What the agent is good for

Two honest uses. First, proposing a layout for a scene you've described in prose - describe the picture, get a plausible set of boxes with names and descriptions, then edit the ones the model got wrong. That's much faster than drawing nine boxes from scratch, and the boxes are the part that takes the time.

Second, tracing a reference photo: with the image wired in, the agent can turn "what's where in this picture" into boxes on the canvas. Useful when you're remixing someone's composition, or rebuilding a shot you liked.

What it won't do is decide your artistic intent. Expect to fix box edges, and expect the descriptions it writes to be bland - they're the same "a red car, side view, occupying the lower third" register that all captioners produce. The prompt you actually want is still yours to write.

Install

Part of the same pack:

comfy node install comfyui-nynxz-nodes

Or search Nynxz in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzNodes
cd ComfyUI-NynxzNodes
pnpm install && pnpm build

Restart. No Python dependencies of its own - and no separate VLM install, either: the encoder comes from ComfyUI's normal CLIPLoader.

Getting the encoder right

The one real trap is picking a text encoder that can't see. A CLIP that generates fine but has no vision tower gives you an agent that has nothing to look at, and the symptom is boxes appearing in places that have nothing to do with your image or your description. If the model dropdown you're using elsewhere for Krea 2 works, use that; if you loaded an old CLIP-L because it was the first thing in the list, that's your problem.

Second trap: the photo arrives in the editor but not in your render. That's normal - image here is for the editor. If you want the model to actually work from that photo, Fresco Inputs' photo is the one that gets edited, and for Layout the photo is a tracing aid, not conditioning.