Layers From Bounding Boxes

Layers From Bounding Boxes is the compositor's import path. It takes an image batch plus a matching set of bounding boxes and turns them into a real layer stack - one layer per frame, each placed by its own box. The official description says it best: use it when a node emits layers as a batch, because a batch carries a single placement for every frame and the individual positions would otherwise be lost.

When would that happen? Real-world example: the Seedream Layer Separation partner node returns a subject cut out against the background as a batch of frames with per-region boxes. Feed that into this node and instead of one flattened result you get separate editable layers sitting where the model detected them. Core object detectors work too - Create Bounding Boxes, RT-DETR, or SAM3 all emit the BOUNDING_BOX type this node eats. It's the bridge between "the model found the objects" and "I can reposition them by hand."

How it works. The mapping is index-aligned: the nth box places the nth frame. Each box's width and height scale the layer to fit it, and frames without a matching box land at the origin. Boxes can arrive as proper BOUNDING_BOX objects, an ARRAY, or a JSON string - the node accepts all three. Two details matter if you're feeding it detector output: normalized elements (a dict with a bbox key in 0–1 coordinates) need canvas_width/canvas_height to resolve to pixels, and if those are 0 the node throws instead of guessing. And when a box carries metadata - name (or desc) becomes the layer name, z_index sets stacking, content_rect defines the real content within a padded frame - the node honors it.

The inputs that matter.

• image and bboxes are both required and must line up index-for-index.
• mask - per-frame transparency, 1 = transparent (the LoadImage convention).
• layers - append onto an existing stack instead of starting fresh.
• crop_to_content - on by default. When a frame has a content_rect, it crops to that rect and places the content at the box position plus the rect offset, so padded frames land at their true spot.
• canvas_width / canvas_height - explicit document size; 0 derives it from the placed layers.

Output. A single LAYERS output, ready for Create Layered Image. From there you can hand-adjust each layer in the editor, exactly as if you'd built the stack with AddLayer.

Getting it. It ships with ComfyUI core - nothing to install, it updates with ComfyUI itself. Like the rest of the compositor, it's experimental and new (landed around August 2026), so treat its behavior as a moving target.

Where people get burned. Index alignment is the whole game: if your boxes and frames fall out of sync, layers land at wrong positions and nothing tells you why. Normalized boxes silently error out if canvas dims aren't set, so when a detector gives you 0–1 coordinates, set canvas_width/canvas_height explicitly rather than leaving them at 0. And remember crop_to_content shifts where content lands - for frames that are already tight, you may want it off to avoid surprise cropping.