LayerUtility: Collage(Advance)

Feed this node a batch of images and it randomly arranges them into a single collage - think a scrapbook-style grid of overlapping, variably-sized photos on a canvas, not a neat evenly-spaced grid. It's a nice-to-have for turning a batch of generations into one shareable sheet, or building reference boards, without doing the layout by hand.

The mechanism leans on randomness by design: images get placed and sized with some jitter rather than uniform tiling, and a seed controls that randomness so you can reproduce a layout you liked or roll for a new one.

Inputs that shape the result:

• images - your batch to collage.
• canvas_width / canvas_height (default 2048×2048, range 512–8192) - output canvas size.
• uniformity (0–1, default 0.5) - how consistent the image sizes are within the collage. Higher values mean less size variation; lower values mean more randomness in how big or small each photo ends up relative to the others.
• border_width and rounded_rect_radius - a border around each photo and how rounded its corners are, purely cosmetic framing choices.
• background_color - the canvas color showing through wherever photos don't cover.
• seed - controls the random placement; re-run with the same seed to get the same layout back.

There's an optional pair worth knowing about if you want smarter placement than pure randomness: florence2_model (optional input) plus object_prompt (default "face"). Wire in a Florence2 model and this node will use it to recognize objects - described in the README as being used "for object recognition and cropping" - meaning it can crop toward the thing you named in object_prompt rather than using the raw source frame. Skip both and the node just uses full images with plain random placement.

Outputs are image (the finished collage) and mask, in case you need to know which regions of the canvas ended up covered by actual photo content versus background.

Installing it: ComfyUI Manager, search "ComfyUI Layer Style Advance," or manually git clone https://github.com/chflame163/ComfyUI_LayerStyle_Advance into custom_nodes, then install_requirements.bat (or install_requirements_aki.bat) or pip install -r requirements.txt, then repair_dependency.bat, then restart ComfyUI. Collage itself has no model download of its own - if you're not using the optional Florence2 input, there's nothing extra to fetch. If you do want the Florence2 object-recognition path, that model downloads automatically on first use, or you can grab it ahead of time from the BaiduNetdisk link in the README and place it in ComfyUI/models/florence2.

One localization note from the README worth passing along: the node's UI supports Chinese, French, Japanese, Korean, and Russian labels (credited to a separate globalization-node-translation project), which doesn't change how the node works but explains why you might see non-English text in a shared workflow screenshot using this node - it's not broken, it's just a different display language.

Where people get surprised: because placement is randomized by design, don't expect a tidy grid - if you want a strict, evenly-spaced layout instead, this isn't the node for it; look for a dedicated image-grid or contact-sheet node elsewhere. And if results look repetitive across runs, check that you're actually changing seed rather than leaving it fixed.