RandomTensor

RandomTensor does one boring, indispensable thing: it generates a random TORCH_TENSOR of a shape you specify. In a pack where every node feeds on tensors - ReshapeTensor, SliceTensor, TensorsToDataset, TrainModel - this is the node you use to test the plumbing before real data exists. Build your training graph, feed it a (64, 784) random tensor, and find out your batch shape is wrong without downloading a single MNIST image.

It's part of EternalKernel PyTorch Nodes (TashaSkyUp), the raw-PyTorch-inside-ComfyUI pack. RandomTensor is squarely in the "debug and prototype" corner of it, which is a legitimate and heavily-used corner.

How it works

You give it a shape string and pick an init_method:

• rand - uniform values in [0, 1). The default.
• randn - standard-normal (bell curve) values.
• randint - random integers from 0 to 99 (the hardcoded range).
• the _like variants (rand_like, randn_like, randint_like) - same draws, but they build a throwaway tensor of your shape first, which is a quirk you can safely ignore.

The dtype dropdown lists every dtype torch knows about (the 55-item enum is auto-generated from the installed torch), defaulting to torch.float32. Just leave it there unless you specifically need ints or a half-precision tensor.

One format gotcha that will bite you on day one: the shape string is parsed by stripping the parentheses and splitting on commas, so it must look like (2,3) - parens included, spaces tolerated. The default (1,1) is your hint. It's a small thing, but this pack's shape-string conventions are inconsistent across nodes (ReshapeTensor wants 1, -1, no parens), so don't assume they match.

Inputs and outputs

• shape (required STRING) - (rows, cols) style, e.g. (16, 10) or (64, 784).
• dtype (optional enum, 55 torch dtypes) - leave on torch.float32 unless you know otherwise.
• init_method (optional enum) - rand / randn / randint / the _like variants.
• Output: TORCH_TENSOR - wire it anywhere a tensor is expected. (2,3) gives you six values; (64,784) gives you a fake batch.

Installing it

Shared with the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TashaSkyUp/EternalKernelPytorchNodes.git
cd EternalKernelPytorchNodes
pip install -r requirements.txt

Restart ComfyUI and it's under ETK/pytorch (or install via Manager, search "EternalKernel PyTorch Nodes").

Troubleshooting

The two failure modes are both easy to read: an Unsupported dtype ValueError means the dropdown served you something the node's internal lookup couldn't match (unlikely but possible across torch versions), and an unparseable shape throws a plain Python int() error - check for missing parentheses. Also remember random values change on every run, which is exactly what you want for testing but a trap if you were hoping for reproducible data. There's no seed input here, so if you need stability, save the tensor with SaveTorchTensor once and load it back.