Flatten Prompt for Caption

Flatten Prompt for Caption takes a prompt full of (emphasis:1.3), (parentheses), and escaped \(weird\) characters, and turns it into clean plain text you'd actually want to save in a caption file. One string in, one string out, no settings - it's the node you reach for when the prompt you generated with is not the prompt you want to train on.

What it actually does

Your working prompt is covered in syntax: (masterpiece:1.2), (detailed face:0.9), a \(cat\) in a park - where the parentheses and colons are instructions to the sampler, not content. Caption files are the opposite: they're training or tagging text where (cat:0.9) is just noise that teaches the model the wrong associations. This node removes the emphasis syntax and weight values, then unescapes the escaped parentheses and backslashes - so the literal \(cat\) becomes (cat) as content, and the weighted (masterpiece:1.2) becomes masterpiece.

Under the hood it parses the string properly rather than doing a dumb regex strip: it tracks parenthesis depth so a real (cat) in the text survives while weighted emphasis gets removed, and it handles backslash escapes so you don't lose a literal character the sampler was treating as punctuation. That's the difference between "clean" and "destroyed."

Where it fits

The pack's caption pipeline is: tagger output or handwritten captions plus prompt text, merged and cleaned, then written as same-name .txt files alongside images (via Image Saver with write_caption, or Caption File Saver). Flatten Prompt for Caption sits at the front of that chain - take the positive prompt that just produced a great image, flatten it, and you have a caption that describes the image without the sampler's annotations. The pack also has sibling caption tools (Merge Caption Tokens, Remove Caption Tokens) to tidy the result further, but flattening the prompt is the step people most often forget.

The one input

Just string - the prompt text. Output is string, the flattened version. No threshold, no mode toggle, no surprises.

The honest caveats

It doesn't know what a "good" caption is - it only removes sampler syntax. Your (masterpiece:1.2), (best quality) flattens to masterpiece, best quality, which is exactly what you want if you're captioning Illustrious-style training data, and exactly what you don't want if you're captioning for a model that shouldn't have quality tags at all. Also, if your prompt uses unusual syntax like numbered steps or Dynamic-Prompts alternatives, the flattening is conservative - it's built for standard emphasis syntax, and fancier constructs will pass through. That's usually the right failure mode.

Installing it

Ships in the ComfyUI-Info-Prompt-Toolkit pack. Via ComfyUI Manager - search "Info Prompt Toolkit" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Restart ComfyUI. No model downloads. Small single-author pack (GPL-3.0), modest footprint - but the niche it fills, prompt-to-caption without the sampler's syntax, is a real gap in the standard install.