Ptdm Icdf

PtdmIcdf is the reverse of PtdmCdf: you hand it a probability q (a number between 0 and 1) and it hands back the value where the cumulative distribution hits that probability. That's the inverse CDF, aka the quantile function - the thing behind every z-score lookup table and confidence interval you've ever seen. If you want "the value such that 95% of draws fall below it," this is the node.

It's part of ComfyUI-Pt-Wrapper, HowToSD's no-code PyTorch environment. It pairs with a Ptd* constructor, and it's one of the more useful query nodes for turning probabilities into concrete numbers in the graph - computing quantiles, defining cutoff thresholds, or reverse-engineering a distribution's behavior.

How it works

The node parses your q with literal_eval, converts it to a float tensor, and calls distribution.icdf(q). That call is deceptively simple because of all the work done elsewhere: PyTorch's Normal and Uniform implement icdf natively, but Gamma, Poisson, and StudentT don't - the pack's Ex subclasses patch those in using scipy's ppf functions. So PtdmIcdf only works as well as it does because the pack filled torch's gaps.

Inputs that matter

• distribution - a PTDISTRIBUTION from a Ptd* node.
• q - the probability, as a text literal between 0 and 1: 0.5 for the median, 0.95 for a 95th percentile. A tuple like (0.025, 0.975) gives you a pair of bounds at once.

Output is a TENSOR with the value(s) at those probabilities. For the median (0.5) of a normal you should get back something very close to loc - a nice self-test.

Installing it

Same as every node in the pack: ComfyUI Manager → search "Pt Wrapper" → install → restart, or:

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

Then restart ComfyUI. The heavy first install (matplotlib, pandas, scipy, scikit-learn, transformers, datasets, pinned gensim) - and scipy is load-bearing here, since the patched icdf implementations for Gamma, Poisson, and Student-t call into it.

Common issues

The classic mistake is feeding q outside [0, 1]. A probability of 1.5 or -0.2 is undefined - some distributions will return inf or nan, others throw. Keep it in range. Also remember q is a text input: type 0.5, not 50%. And if a distribution you're using throws NotImplementedError on icdf, you've hit one the pack didn't patch - stick to the Ptd* constructors and you're safe.