Extract PDF (PyMuPDF)

This is the kitchen sink node of the pack: give it a PDF (from Load PDF (PyMuPDF)) and it pulls out the embedded images, the text, and per-item metadata in a single run. If your document pipeline needs both "what's in this PDF visually" and "what does it say," this is the one to reach for - the more focused text-only and render-only siblings exist for when you don't want to pay for both.

How it works

Under the hood it's straight PyMuPDF. For every page it runs page.get_images(full=True) to list the images actually embedded in the file and turns each one into a tensor via fitz.Pixmap (routed through a PyMuPDF RGB conversion when the source has an alpha channel, so the batch stays clean RGB). Text comes from page.get_text("text"), chunked per page with a --- filename page N --- header so you can tell where each piece came from. It then stacks everything into a single IMAGE batch and returns the text as one big STRING.

The knobs that matter:

• extract_images / extract_text - flip one off and you skip half the work. For a text-heavy document you don't need images; for an image scan you don't need text.
• max_pages and max_images - the guardrails. PDFs can be enormous, and the code is honest about it: max_pages caps how many pages get looked at, max_images caps how many images get pulled total. Set them to 0 for no limit, but know that a 400-page magazine will happily try to extract hundreds of images into one tensor batch and run you out of RAM.
• render_pages_if_no_images - the clever fallback. When a page has no embedded images (common with scanned or vector-heavy PDFs), this renders that page to an image at render_dpi instead, so you still get something visual out.
• resize_mode - fit_first sizes everything to the first image's dimensions, fit_smallest to the smallest, pad_largest zero-pads the smaller ones to the biggest canvas. This exists because a tensor batch must be a rectangle; the default fit_first is usually fine.

What comes out

Five outputs, and they map to the three things you'll wire:

• images (IMAGE) - into an img2img or ControlNet pipeline, or just a preview.
• text (STRING) - straight into a text display node, an LLM node for document Q&A, or Save PDF Text (PyMuPDF) for a file on disk.
• metadata_json (STRING) - per-image info: which PDF, which page, whether it was an embedded image or a page render, and the DPI.
• page_count and image_count (INT) - handy for routing or for a text overlay.

Install & gotchas

Same story as the rest of the pack: ComfyUI Manager or clone + pip install -r requirements.txt (PyMuPDF only), then restart. If the node doesn't show up, PyMuPDF isn't importable in ComfyUI's Python environment - install pymupdf (not fitz, there's a stale decoy package by that name) into the environment ComfyUI actually runs from.

The one thing that reliably trips people: batch mode. Load PDF (PyMuPDF) set to batch hands this node all the PDFs in the input folder, and the two limits behave differently. max_pages is applied per document - each PDF in the batch gets its own 20-page cap. max_images, by contrast, is a global accumulator: once 20 images have been pulled across all files, the whole run stops, so an image-heavy folder can get you 20 images total even if you expected 20 per doc. Set both knowing that.