Meta Prompt Extractor

Every ComfyUI PNG is secretly a project file. The workflow that made it - node graph, seed, prompt - is embedded in the file's metadata as text chunks, and this node exists to crack that open and hand you the good parts: the positive and negative prompts, the image itself, its dimensions, its path. It's the "I found this image, I want its prompt" node, pointed at a file that lives anywhere on disk, with a full file browser attached.

Point it at any PNG, JPG, JPEG, WebP, or JSON file and it extracts:

• positive_prompt / negative_prompt - the generation text, from either a ComfyUI workflow, an A1111-style parameters string, or a workflow JSON.
• image - the image as a tensor, so you can feed it straight into img2img or a VAE.
• mask - loaded from a matching <image>_mask.png file, or painted in with the built-in Mask Editor.
• path, width, height, resolution - the bookkeeping.

One output node feeding an img2img pipeline with both the picture and the prompt that made it - that's the common setup.

How it finds the prompt

Straight metadata read only gets you so far. If the workflow chunk is there, it parses it; if the file carries the A1111 flat parameters block, it reads that dialect too. But workflows get complicated - prompts built by wildcard nodes, chained text, custom encoders - so when there's no clean prompt string, the node walks the graph: it finds the sampler as the anchor, follows connections backward through intermediate nodes, checks a registry of known custom nodes, and finally falls back to a smart scan that returns the most natural-language-looking result. That's the part the author is most proud of, and it's genuinely useful on messy real-world files.

The one thing it can't do is invent what was stripped. If a PNG was re-encoded by an image host, the chunks are gone and no extraction brings them back - the pixels never had the text. Keep the originals.

The inputs that matter

• image - an absolute file path. There's a Browse button that opens the full filesystem browser: breadcrumbs, search, sorting by name/date/size/dimensions, favorites, multi-select. You can also just drag a file from your OS onto the node.
• use_conditioning - flip this on and the optional conditioning / conditioning_negative inputs take priority over file extraction. Connect a CLIPTextEncode and it auto-turns on. This is the escape hatch when the file's metadata is useless but you have live conditioning you'd rather pass through.
• batch_folder - a folder path; the node loads every supported image in it in one pass.
• batch_file_list - a JSON array of exact file paths, when you want a specific subset rather than a whole folder.

Installation

It ships in the ComfyUI-ialhabbal suite, so install once and it's there. ComfyUI Manager search ComfyUI-ialhabbal, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ialhabbal/ComfyUI-ialhabbal.git

Restart, then find Meta Prompt Extractor under the utils category. It uses standard libraries - no extra model downloads.

Common issues

• Empty prompt from a PNG - check where the file came from. Hosts like Reddit recompress and strip chunks; only files that kept their original bytes still have the workflow.
• Won't open my folder - the browser can navigate any drive, but watch permission issues on network or system folders.
• Prompt comes out garbled - you're probably hitting a wildcard or LLM-built workflow where the graph walk landed on a fragment. Try the conditioning override with a live CLIPTextEncode instead.
• Batch returns nothing - double-check the paths in batch_file_list are valid JSON and the folder in batch_folder actually contains supported extensions.

For a utility node it punches above its weight: prompt recovery, batch metadata mining, and a img2img starter all in one.