tensor shape

tensor shape is one of those nodes whose name sounds obvious and is exactly backwards. It does not look at an existing tensor and tell you its shape - that's its sibling get tensor shape from the same pack. This one constructs a shape tuple out of the five integer inputs you type in, so you can hand it to a node that wants a TUPLE of dimensions. Think of it as a "specify a tensor's dimensions" node rather than an inspector.

You'll reach for it inside 3D and batch workflows when some node needs to know how big a tensor should be before it exists. The defaults tell the whole story: Dim1=1, Dim2=512, Dim3=512, Dim4=3, Dim5=0. Zero means "stop here", so the defaults assemble into (1, 512, 512, 3) - a single 512×512 RGB image, i.e. exactly the layout ComfyUI's image tensors use. Want a 2D matrix of 64 floats? Set Dim1=64, Dim2=64, and everything after to 0. The rule is simple: the first Dim you set to zero is where the tuple ends.

In this pack the shape you build feeds tensor new, which fills a tensor of that shape with either a constant value or random noise. That's the actual pipeline this node belongs to: shape → tensor → image. If you're coming from pure 2D ComfyUI, this is the moment you've crossed into a graph that talks in raw torch tensors instead of prepackaged IMAGE objects.

Inputs worth your attention, honestly, are just the five Dim1–Dim5 integers - and only the ones you care about, since trailing zeros get trimmed. There's exactly one output, tensor_shape (a TUPLE), and it wires into anything that accepts a shape tuple.

One thing to know before you rely on it: this whole pack is a one-developer utility collection, and it leans on the Tensor/TUPLE types and the general conventions of ComfyUI-3D-Pack, so if a workflow built around these tensor nodes won't load, check whether that 3D pack is installed first. The author also renames nodes between releases, and old ones go red - recreate the node rather than hunting for a missing one.

Install it the usual way: ComfyUI Manager → search "ComfyUI-3D-MeshTool" → install, then restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/807502278/ComfyUI-3D-MeshTool
cd ComfyUI-3D-MeshTool
pip install -r requirements.txt   # kiui, xatlas, pyfqmr, plyfile

Nothing here needs a model download - the tensor nodes are pure math on numbers you type. If you mainly work with regular images, this node is marginal; if you're feeding a mesh or multi-view render pipeline that wants precise tensor dimensions, it's the cleanest way to spell them out.