Extract images from Word documents

Extract images from Word documents without flattening them

A modern Word document is a ZIP package of XML parts and media files. This extractor follows the relationships from the main document, headers, footers, footnotes and endnotes to the image entries they actually use. That matters because simply copying everything under word/media can include an orphan while missing the fact that one image is reused many times.

Each embedded package entry is kept byte for byte. The page calculates SHA-256 from those bytes and records every occurrence separately. If two different package paths contain identical bytes, the manifest marks the duplicate without throwing either path away. Externally linked images are reported as links, not presented as missing downloads: their bytes were never embedded in the DOCX.

“Original” means the bytes stored inside the DOCX

Word may resize, recompress or otherwise transform a picture before saving it. A DOCX rarely retains the camera filename, and it cannot restore a design file that was replaced before the document was saved. Here, original has a narrow verifiable meaning: the exact package entry recovered from this document. Its hash in the manifest must match that entry after the exported ZIP is reopened.

PNG, JPEG, GIF, WebP and other recognizable raster formats are identified from their signatures rather than trusted extensions. Browser-unsupported EMF, WMF, TIFF and SVG entries remain in originals/ even when the workbench cannot rasterize them. Keeping an unfamiliar original is more honest than silently omitting it.

Word crops are evidence, not a rendering promise

A picture cropped in Word normally retains hidden pixels in its package file. The inventory displays that a crop exists and flags the image for review. It does not claim to reconstruct every transform, effect, rotation, group, shape mask or version-specific Word rendering detail. The resize step starts from the stored image and asks you to make a new explicit crop.

Choose Fill and crop when every file must cover the full target rectangle. Choose Fit with padding when every source pixel matters. Neither mode stretches the picture. Upscaling is off by default, and a small source remains small inside the chosen canvas until you explicitly allow enlargement.

Continue to the same bulk image resizer without uploading again

After inventory, the extracted bytes are already in the shared workbench used by the bulk image resizer. Set any whole-pixel width and height, choose the output format, name pattern and padding, then inspect the flagged cards. Large crop loss, alpha, animation, Word crop metadata, a source smaller than the target and an unsupported decoder all produce visible reasons for review.

The batch rule handles ordinary files together. Individual horizontal and vertical focus controls are available where a centered crop would cut off the subject. The output can be PNG, JPEG or WebP. JPEG cannot carry transparency, so a transparent choice becomes white rather than an accidental black background.

The download is verified before it is offered

The result contains originals/, a folder named for the requested dimensions, and both JSON and UTF-8 CSV manifests. Processed images are decoded and encoded in a local worker. Each successful file is inspected again and must report the chosen width and height. The ZIP is then reopened, every entry is read, and every byte sequence is hashed against the entry that was meant to be written.

Malformed archives, more than 20,000 package entries, a package entry above 512 MB, total declared expansion above 1 GB, extreme compression ratios and actual expansion beyond those declarations fail with a specific package-level message. Those structural limits defend the browser from a ZIP bomb; they are not a file-upload quota. The local-processing guide explains how to watch the network while the document is handled. If the source is a web page rather than Word, use the website image extractor instead.