kubakub scene render

This is the node the 3D half of the pack is built on. You give it the architect's model, it gives you the projector's view of the facade plus a machine-readable description of every surface in it - so "give every window its own prompt" stops being a masking exercise and becomes a checkbox.

kubakub scene render accepts .blend, .fbx, .obj, .abc, .glb, .gltf, .stl, .ply and USD, runs Blender 4.x headless as a subprocess, and returns a clay render, metric depth, normals, and a folder of ID passes that kubakub regions from id maps already knows how to read.

How it works, because the mechanism is the good part

Blender flattens everything render-visible into world-space meshes (modifiers applied, instances made real), renders a Workbench clay view, and takes a single Cycles sample without AA that stores, per pixel, which face it sees plus that face's world position and normal. The ID passes themselves are then computed in numpy from that buffer - which is why changing a threshold doesn't trigger another render. The pack's test facade is 20k faces at 3200×2160 and takes about 4.5 s on the GPU.

The result is cached per file, camera, size and frame in ComfyUI/temp/kuba_scene3d.

Which passes to ask for

passes is one space-separated line, and the defaults are a good starting set:

• shelves - the facade's depth steps, found from the depth histogram along the facade normal and named by their offset from the main wall. shelf_+0.35m means 35 cm proud of the wall. shelf_min_step_m decides how close two steps can be and still count separately.
• layers - fixed depth bands around the main plane (deep_recess / recess / facade / relief / projection), set by layer_bands_m as four limits in metres, + toward the camera.
• facing - front / top / underside / side / back relative to the facade normal. This is how you select every upward-facing ledge at once.
• planes - connected coplanar patches, tuned by plane_angle_deg and plane_offset_m.
• parts - loose parts: every separately modelled stone, frame, cornice.
• elements, sections, floors - the architecture passes, and the most useful of the lot. elements splits the facade into windows (including blind windows, with arcades split per arch), window_frames, columns, cornices, other relief, wall and roof. sections gives left / centre / right where the wall actually steps forward or back. floors gives ground_floor, floor_1, floor_2, roof, derived from the rows of windows and the cornice belts.
• objects / materials / collections - straight from the file, written only when there's more than one.

Names from these passes are group and tag fodder downstream: group:windows, tag:left, tag:floor_1 group:windows. Grouping "*" splits every window and column into its own region, grouped by class. On the test facade, parts as regions gives 423 regions - every stone, voussoir, frame and pediment.

preview_pass chooses which map shows on ids_preview (default shelves). min_size_m (default 0.5) merges pieces smaller than that across the wall into whichever neighbour they border most - the crowd can't read them, and 3018 plane patches become 1352 at 0.5 m. Set it to 0 if you're doing fabrication work and want every stone.

Camera, units and outputs

file empty means the built-in sample building, so the node runs as-is. For your own file, paste the path - right-click in Explorer, Copy as path, quotes and all.

camera takes an object name; otherwise the file's active camera, else its first camera, else a front camera framing the model (it finds the facade side by itself - a facade is wide and shallow). A connected projector overrides all of that and places the virtual projector in real metres. Set width / height to match; 0 means "the file's render resolution". FBX, OBJ and Alembic carry no resolution, so the node looks for a size in the camera name (that's why you'll see names like RENDER_CAM_3200_2160PX in these pipelines) and otherwise defaults to 1920×1080 with a note. FBX round-trips like to flip lens shift; when the imported camera sees almost none of the model and the flipped shift frames it, the node uses the flipped one and tells you.

unit_scale multiplies the model into metres - 0.01 for a centimetre file. This is not optional housekeeping: some FBX exports arrive about 0.5 m wide, and the report warns when a model is under 2 m or over 2 km. Trust that warning; every distance downstream depends on it.

Ten outputs, and they all earn their place: clay (the projection view), ids_preview, depth (near = bright), normal, coverage (1 where the model is - a scope for the region nodes), folder (the ID-map folder, into kubakub regions from id maps), scene_json (camera matrices, resolution, names, main plane, shelves), report, scene (into scene measure / preview / relight / pieces), and depth_gray - linear depth from near = white, which is what you feed to a depth-to-relief node if you're going as far as milling a physical panel.

save_folder copies the ID maps, legends, clay and scene.json out to somewhere permanent, in the format Houdini and After Effects tools read. Worth doing at the start of a project rather than re-deriving passes later.

Install

cd ComfyUI/custom_nodes && git clone https://github.com/kubakubkub/ComfyUI_KubaNodes
python -m pip install -r ComfyUI_KubaNodes/requirements.txt

Then install Blender 4.x. The pack finds it in C:\Program Files\Blender Foundation or via blender_path / blender = ... in kubakub.ini. Blender is never imported into ComfyUI's Python - it's a subprocess - and this is the design decision that makes the 3D half of this pack installable at all. Compare it to the CUDA-extension node packs that need matching PyTorch, Visual Studio build tools and a pinned wheel to even import.

Two limits worth knowing before you plan around it. .c4d can't be read outside Cinema 4D - export FBX or Alembic. And the source file is only ever read, never saved; if your scene relies on modifiers or simulations, they're evaluated by that Blender version, which is why the pack is tested against 4.5 specifically.

If the node errors, the report is usually explicit: it tells you the faces, the camera chosen, the depth range, the shelves found, which passes were written, whether it hit the cache, and where the ID-map folder is. Nine times out of ten the answer is in there.