PNG product plates look harmless until a marketplace scraper flags an AI training tag, a generator comment, or a SoftBank-style label buried in textual chunks. JPEG teams already have EXIF muscle memory. PNG teams often assume "no EXIF means clean," which is wrong. This note is the intake checklist I use when a seller ships PNG masters from Midjourney upscales, CapCut still exports, or design tools that stamp provenance into tEXt / iTXt / zTXt chunks.
PNG does not carry classic EXIF the same way JPEG does, but it happily stores free-form text chunks, XMP packets, and AI-oriented comments. A cleaner that only nukes APP1 EXIF can leave those chunks intact. Marketplace review bots that parse PNG text will still see the generator string. Treat PNG as its own pipeline, not as "JPEG without EXIF."
Ask the seller three things in writing: which tool produced the plate, whether the marketplace listing claims human photography, and whether a sealed AI-provenance archive must be kept for their own compliance. If the listing claims photographic originality while the file still carries generator metadata, that is a policy problem, not a pixels problem. Escalate before stripping.
Create sealed_png_ai_meta and delivery_png_clean. Never overwrite sealed. Copy into delivery, strip on the copy, then verify. If a later dispute asks what the original file contained, you need the sealed bytes, not a story.
Usually strip: generator comments, AI training labels, tool fingerprints that identify a model name, and redundant XMP bags the portal rejects. Usually keep: color profile if the catalog depends on it, and any rights or caption fields the seller still wants on the public asset. Ambiguous "clean everything" tickets destroy useful IPTC-like text. Name the artifact class in the ticket: AI-only metadata versus full privacy scrub.
When the brief is specifically AI labels and related PNG text bags rather than a total privacy wipe, I use a dedicated page instead of a blunt EXIF hammer. Bookmark: png AI metadata remover. Run it on delivery copies only. Spot-check one file in a hex-aware viewer or a metadata inspector for leftover generator strings. Spot-check that the image still opens and that colors did not shift in an unexpected way.
For batches under twenty, check every file. For larger batches, check five random plus the first and last. Search for obvious generator tokens in text chunks. If any sample fails, halt and restore from sealed. Shipping a half-clean folder is worse than shipping late.
Some marketplaces only complain after a competitor report. Others reject at upload. Reproduce the rejection on a single PNG before you strip a whole drop. If the portal accepts AI labels and the seller discloses generative assist, do not strip "just in case." Extra stripping is not professionalism when disclosure is required.
Visible watermarks, SynthID-class signals, and textual PNG AI tags are different problems. This desk note only covers metadata and label bags a browser tool can strip locally. Do not promise invisible watermark erasure. Do not promise platform labels will disappear from a social feed after upload.
In the delivery note list: sealed path, delivery path, which AI tags were targeted, which sample reproduced the portal issue, and which tool page was used. Include the date and operator initials. Three months later someone will ask why the PNG no longer mentions a generator; answer from the note, not from memory.
PNG AI metadata is an operations problem with a narrow mechanical step. Keep sealed archives, strip only what the brief names, verify chunks are gone, and leave a paper trail. That is enough to keep marketplace uploads boring in the good way.
When a seller sends both formats, run JPEG and PNG cleaners as separate passes with separate logs. A shared "metadata remover" button that treats every container identically will miss PNG text bags or over-delete JPEG captions. Split the queue, split the verification, merge only at the zip step.