CV Parse Object Transform

The pack ships a 3D viewer with a move/rotate/scale gizmo (CV 3D Object Transform), and dragging a model into place is by far the fastest way to decide where it should go. The problem is that a nudge inside a viewer is a widget value, and widgets don't wire into nodes. CV Parse Object Transform is the node that turns that widget into a 4x4 matrix and three vectors, so a placement you found by eye becomes data - the mesh, the annotation anchors and the silhouette all follow the model you parked.

That's the standard ComfyUI value-node pattern, applied to a 3D transform: one authoritative source, many consumers.

What you feed it

object_transform is the JSON string the viewer writes - position, quaternion [x, y, z, w], scale, in Three.js world space. You can wire it from the viewer or just paste it. Scale may be a plain number, missing keys default to identity, and an empty string is identity and counts as valid - that's what an untouched widget holds, so a fresh node doesn't look broken.

space picks the coordinate system for all the outputs: OpenCV (Y down, Z into the scene) or the Three.js convention. Match this to CV Mesh From 3D Model's axis_convention, or the transform you get back won't apply to the mesh you got back - the anchors end up mirrored through the object and you'll spend twenty minutes wondering why. The input string is always Three.js regardless of what you pick here.

What comes out

Five outputs, all in the selected space:

• matrix - the 4x4 float64 model matrix, ready for CV Transform Points 3D.
• position - 1-D [x, y, z] translation.
• quaternion - 1-D unit quaternion [x, y, z, w]. In OpenCV space it's conjugated by the axis flip, so it still recomposes matrix together with position and scale.
• scale - per-axis [x, y, z]. The gizmo can scale non-uniformly, so this is a vector even though the transform node only emits uniform scales.
• valid - false when the string couldn't be read, with identity on everything else.

valid is the one to actually wire up. A half-typed JSON string returns identity rather than halting the run - deliberately failure-tolerant, and a silent trap if you don't branch on it: your render quietly puts the model at the origin. Send valid into an If/Else Switch and fall back to a preset placement.

One subtlety worth knowing: what comes back is the whole placement as a single rotation. If the transform node was set with a model_up option, this parses to the combined turn - not to that dropdown plus the rotate_* values separately. So don't try to decompose it back into the widgets that produced it.

Install

Manager → search ComfyUI CV, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv
cd comfyui_cv && pip install "opencv-contrib-python-headless~=5.0.0.93"

Restart. Python ≥ 3.12, ComfyUI on the V3 node API. This node is pure Python and JSON - no cv2 class, no model download, no contrib dependency - but the pack it ships in wants the pinned contrib headless wheel.

Common issues

• valid is false and the model sits at the origin - the widget string is malformed or half-edited. Branch on valid, and check the string is real JSON before blaming the parse.
• The model appears mirrored or rotated 180° - space and the mesh's axis_convention disagree. This is the one mistake this node makes easy, so make it the first thing you check.
• Anchors don't ride the object - the anchors have to come from the same model file as the mesh. Different GLB, different geometry, same visual placement.
• Scene 3D nodes going red - some 3D workflows in the pack depend on other packs (Inspire Pack, Custom-Scripts, Basic Data Handling). Install those from Manager if a workflow from the repo complains.

The pack's README is explicit that this is a personal, LLM-assisted project without production support - worth remembering when a viewer widget is the only record of a placement you spent time finding. Copy the JSON somewhere safe.