JPEG vs PNG for Provenance: Where Assertions and Metadata Live in Each Format
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
JPEG carries XMP in application segments that a re-encode discards; PNG carries text chunks that converters drop silently; both can hold EXIF-style data. ProvenanceGuard inspects and embeds in both formats, up to 25 MB, reading all of those territories. Choose the format for the image's purpose, and let inspection, not assumption, tell you what a given file actually contains.
Arguments about which format is 'better for provenance' usually end when someone inspects real files: both formats lose everything in a platform re-encode, and both carry a record perfectly well in direct delivery.
Pipelines break provenance in boring places: a design tool exporting without metadata, a CDN recompressing on the fly, a social scheduler re-encoding before upload, an archive converting everything to one format. Each is a transform, and each transform is testable. The lineage tool's local transforms and the vault battery's transforms exist precisely so a team can run its own pipeline's steps against a sealed sample and watch which step removes what, instead of discovering it from a published asset that no longer verifies.
What is format anatomy? the internal structure of an image format that decides where metadata can live: JPEG's application segments for XMP, PNG's named text chunks, and each format's EXIF conventions, and what re-encoding does to each.
The problem behind JPEG and PNG provenance
Format details decide more provenance outcomes than policy documents do. JPEG carries XMP in application segments that re-encoding discards; PNG carries text chunks that many converters drop silently; EXIF, written by cameras, is routinely stripped by platforms and says little about AI generation in any case. ProvenanceGuard works with JPEG and PNG because their structures are where its inspection, XMP embedding and fingerprinting are defined and tested, and it inspects all three metadata territories, XMP, EXIF and PNG chunks, rather than assuming any one of them is where the truth lives.
That byte-level honesty changes decisions. A file that arrives with a valid C2PA manifest needs different handling from a file whose metadata was already stripped by an earlier platform hop, and both differ from a file that was never marked at all. Checking first stops a common mistake: sealing and republishing an image while assuming its original manifest is intact, when in fact the copy in hand lost it two uploads ago. The inspection result becomes the baseline the rest of the record is built on.
How ProvenanceGuard handles it
Inspection also has limits worth stating. A marker being present says the bytes contain it; it does not say the claims inside it are true, and a signature check says whether a ProvenanceGuard assertion was altered, not whether the disclosure text a creator typed was accurate. Provenance records what was declared and when it was registered. Treating it as more than that, in either direction, is how teams end up both over-trusting and under-using the records they have.
Only JPEG and PNG files are supported, up to 25 MB, because those are the formats whose byte structures the inspection, embedding and fingerprinting are built and tested against. A file in another format is refused rather than guessed at. That refusal is a feature: a provenance tool that silently approximated on formats it does not understand would produce records that look solid and are not, which is worse than no record at all.
What survives distribution, and what does not
Inspection comes before sealing, because an asset may already carry signals. ProvenanceGuard reads the actual bytes of an uploaded JPEG or PNG and reports what is really there: C2PA and JUMBF markers, an XMP packet, EXIF data, PNG text chunks, any AI disclosure already present in the metadata, and any prior ProvenanceGuard assertion, including whether that assertion's signature still verifies. Nothing is inferred from the filename or the uploader's claims; every field on the inspection result is derived from the file itself.
What sealing does not do is just as clearly bounded. It does not certify that a disclosure statement is legally sufficient, it does not turn the assertion into a C2PA manifest, and it does not store the uploaded original: the registry keeps the sealed and labelled copies and the record, so verification and downloads keep working, while the original bytes are not retained. The seal is evidence of a declaration made at a time, about a specific file, and that precision is what makes it useful later.
A routine that survives busy weeks
Sealing is the moment a record is created. ProvenanceGuard takes the creator and disclosure the uploader declares, the SHA-256 of the original bytes, and the registration time, and signs that assertion with HMAC-SHA256 under the product signing key. The signed assertion is embedded into the file as XMP, producing the sealed copy. Because the signature covers the assertion's contents, altering the creator, the disclosure or the hash inside the file afterwards invalidates the signature, and the verification page shows that plainly.
Alongside the sealed copy, sealing produces a visibly labelled copy: the same image with a parchment custody strip in the pixels naming it as AI-generated content whose provenance is registered with ProvenanceGuard. The distinction matters. Embedded assertions are machine-readable and fragile in transport; a visible label is human-readable and survives anything that keeps the strip in frame, including a screenshot. Publishing the labelled copy where the audience actually sees the image, and keeping the sealed copy as the record, uses each layer for what it is good at.
Five actions to take
- Inspect one real file in each format before writing your format policy
- Treat social exports of either format as registry-backed, never metadata-backed
- Re-check one published asset through the Recovery lookup so you know the stripped-copy path works before you need it
- Seal the final export of the asset, with the creator and disclosure declared in your own accurate words
- Download and keep both copies the seal produces: the sealed copy as the record, the labelled copy for publication
A worked example: A team standardising its AI asset deliveries
It kept masters as PNG for direct client delivery, where chunks and XMP survive intact, and treated every JPEG social export as registry-backed only; inspection reports, not format folklore, set the policy for its 140 assets.
The lesson generalises beyond this case. Records made at publication time are cheap, calm and complete; records reconstructed under pressure are expensive, partial and contested. Everything in this guide is in service of moving work from the second category into the first.
FAQ
Does converting PNG to JPEG destroy the record?
It destroys the embedded copy of it, as any re-encode does, while the registry record is untouched and the converted file's pixels still fingerprint-match. Converting a master is also exactly the kind of step lineage exists to record.
Can anyone verify a record, or only the owner?
Anyone. Verification pages are public by design, the single-hash lookup is a public API, and an exhibit ID or a file hash is enough. A record only its owner could check would not settle a dispute with a stranger.
Is ProvenanceGuard's assertion a C2PA manifest?
No. It is ProvenanceGuard's own signed assertion, embedded as XMP and HMAC-signed, and labelled that way everywhere. The inspection detects genuine C2PA markers when a file carries them, but the two records are never conflated.
What formats and sizes are supported?
JPEG and PNG, up to 25 MB per file. Those are the formats the inspection, XMP embedding and fingerprinting are built and tested against; other formats are refused rather than approximated.