Frakturpad

Somewhere upstream, something is filtering your prompt text. It might be a content filter on a hosted frontend, a moderation layer on an API, or a naive word-blocklist in a tool that copies your prompt around. Frakturpad is the low-tech, zero-infrastructure answer: it rewrites your text using Unicode Mathematical Bold Fraktur characters - the 𝕬𝖇𝖈 style you've seen on forum sigs - and pads it with invisible word-joiner characters. The words are still perfectly readable to a human and to any model that renders text. The bytes, though, look like garbage to anything doing string matching. The author's README states the intent flatly: "useful for bypassing text filters while maintaining readability."

How it works

The mechanism is two steps, both in the source:

1. Every A–Z and a–z is mapped into the Unicode Mathematical Bold Fraktur block - uppercase starting at U+1D56C, lowercase at U+1D586. So "cat" becomes "𝖈𝖆𝖙".
2. The result is padded with the word joiner character U+2060 between every adjacent character and at both ends.

That second step matters: U+2060 is a zero-width character that tells word processors not to break a line there. Visually it's nothing; structurally it means no two ordinary letters are ever adjacent, which is exactly what defeats a filter that searches for literal words.

The node takes one multiline text input and outputs fraktur_text. That's the whole surface - paste your prompt in, wire the output wherever the prompt string goes.

Where it works and where it doesn't

The README's phrase "maintaining readability" is the key. If your goal is a prompt-side filter bypass so the string survives to the generator, this is genuinely effective - there's a whole ecosystem of text-obfuscation tricks like this, and bold fraktur is one of the more legible ones. If your goal is to generate an image containing that text, be more skeptical: image models that render text don't all know fraktur glyphs well, and you may get garbled letterforms in the output. Know which fight you're in before you reach for it.

One more gotcha: zero-width characters like U+2060 get stripped by some copy-paste paths and a few API layers. If the filter you're dodging lives downstream of a text box that normalizes unicode, the padding may be silently removed and you're back to plain words.

Install

ComfyUI Manager → search ComfyUI_SamplingUtils, or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI_SamplingUtils

Then restart ComfyUI. The real requirements.txt pulls kornia, scipy, pilgram, opencv-python, and unifiedefficientloader>=0.5.0 (the README's list is shorter than what actually installs). And since the README declares the pack DEPRECATED in favor of ComfyUI-UtilsCollection, keep an eye out for where these text nodes land next.