📝 LMM Extract Prompts

If you've ever looked at a folder of PNGs and wanted to know what each one was generated with, this is the node that turns that pile into usable text. The 📝 LMM Extract Prompts node talks to Local Media Manager - Firetheft's ComfyUI extension that lists the images in a directory - takes its paths output, and pulls the positive prompt, negative prompt, and generation info out of each image's embedded metadata. It's a batch metadata scraper with one specific, sensible rule set.

How it works

The only input is paths (LMM_ALL_PATHS), which you connect from Local Media Manager's paths output - the node won't work standalone, and that's by design; it's an adapter for LMM, not a file browser. For each image in the list it reads the metadata, and the interesting part is how it splits it:

• positive_prompt - the positive prompt with every <lora:...> tag removed.
• negative_prompt - same treatment.
• info - the inverse: it keeps only the <lora:...> tags plus everything from Steps: onward (sampler, steps, cfg, seed, model, hashes...). That's the Civitai-style generation block, LoRAs included.

So you get clean, reusable prompts on one side and the machine-parseable generation record on the other - a deliberate split, not an accident. Multi-selection works: select several images and the prompts concatenate across them (non-image items are ignored).

When you'd use it

The classic flow is reconstruction-after-the-fact: you've got a batch of images (someone else's or your own from before you started tracking workflows) and you want to feed the recovered prompts back into a generation or LoRA-training session. Pair the positive_prompt/negative_prompt outputs with a text-input node, and wire info into a metadata saver if you want to re-embed the generation block.

The parsing engine is also robust in a way worth noting: it handles both the A1111-style parameters text chunk and ComfyUI's embedded JSON workflow (it walks the graph to find the CLIPTextEncode nodes that feed a sampler), so it copes with images from both lineages.

Caveats

• It's only as good as the metadata in the files. Screenshots, re-encoded JPEGs, and web downloads have nothing to read - you'll get empty strings, not errors.
• This node requires the Local Media Manager pack installed and in your graph; it's an integration piece, not a standalone prompt reader. If you don't use LMM, you're better off with a plain metadata loader.
• The LoRA-stripping is regex-based on the <lora:...> pattern, so nonstandard tag formats may survive into the prompt output.

Install

Install Light-x02 Nodes plus ComfyUI_Local_Media_Manager (both via ComfyUI Manager, or git clone into custom_nodes):

cd ComfyUI/custom_nodes
git clone https://github.com/Light-x02/ComfyUI-Lightx02-Nodes
git clone https://github.com/Firetheft/ComfyUI_Local_Media_Manager

Restart ComfyUI. The Light-x02 side has no extra dependencies beyond Pillow, which ComfyUI already ships.