Ptdm Log Prob Tensor

The scalar PtdmLogProb scores one value. PtdmLogProbTensor scores a whole tensor of values against your distribution in a single call - the version you actually want when you've sampled a batch and need the log-likelihood of every draw, or when you're feeding a loss computation. One wire, one call, whole batch handled.

It's part of ComfyUI-Pt-Wrapper, HowToSD's no-code PyTorch toolkit. It's the vectorized sibling in the Ptdm* family, and it's the one that shows up in real training workflows - computing negative log-likelihood over a batch of predictions is exactly the kind of thing you'd wire into the pack's loss machinery.

How it works

No parsing, no dtype sniffing - it calls distribution.log_prob(tens) directly and returns the result. The distribution handles whatever dtype your tensor carries, and the output keeps the input's shape. It's about as thin a wrapper as this pack ships; all the real behavior lives in the distribution you built with a Ptd* node.

Inputs that matter

• distribution - a PTDISTRIBUTION from a Ptd* constructor.
• tens - a TENSOR of values to score.

Output is a TENSOR of the same shape, filled with log-probabilities. Sum it with the pack's tensor-math nodes and you've got a total log-likelihood for a whole dataset - a genuinely useful statistic you'd otherwise have to code.

Installing it

Standard pack install: ComfyUI Manager → "Pt Wrapper" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/HowToSD/ComfyUI-Pt-Wrapper

Restart ComfyUI after. First install is slow because the pack pulls a big stack (matplotlib, pandas, scipy, scikit-learn, transformers, datasets, pinned gensim), but this node needs nothing extra.

Common issues

The outputs are logs, so expect negative values and -inf wherever the distribution assigns zero probability - that's the feature, not a bug. If your tensor mixes dtypes, cast it with the pack's PtToFloat32 or similar before scoring; a discrete distribution scored on a float tensor is a mismatch waiting to bite. And as always in this family, if you see NotImplementedError, the distribution you're using wasn't one of the pack's patched constructors.