The pixels change
Published ; updated
A separate registered experiment compared Adobe's Rust 0.5.0 and Python TrustMark implementations on twelve owned images. C2PA signing and removal preserved every decoder outcome. The explicit Rust pixel remover changed the images and left zero exact-payload recoveries for either matching decoder across the 24 marked outputs. Those results concern the tested Q models and public payloads.
Two decoders
Each input was encoded once with Python and once with Rust, using a new registered 68-bit payload, Q, BCH_4 and strength 0.95. The decoders did not agree on every marked image.
Scroll horizontally to see all columns
| Encoder | Decoder | Marked | C2PA signed | C2PA removed | Pixel removal |
|---|---|---|---|---|---|
| Python | Python | 11/12 | 11/12 | 11/12 | 0/12 |
| Python | Rust | 10/12 | 10/12 | 10/12 | 0/12 |
| Rust | Python | 10/12 | 10/12 | 10/12 | 0/12 |
| Rust | Rust | 10/12 | 10/12 | 10/12 | 0/12 |
The Python decoder returned zero positive flags in twelve unmarked originals. Rust returned one, on the unmarked tiles-png control, with a different decoded payload. A positive flag and recovering the expected payload are different observations. Twelve controls supply no population false-positive estimate.
Metadata stays
All 24 files cleaned of C2PA were byte-identical to their marked inputs. Signing and cleaning preserved decoded RGB pixels too. The independent C2PA oracle checked the signed records and found no claim after removal. The development certificate was self-signed and outside public trust lists.
The earlier TrustMark experiment remains ten of twelve exact recoveries before and after C2PA removal, including both misses. This comparison uses different payloads and strength, so the two runs do not establish an improvement in recovery.
PNG attempts
The first runner supplied a JPEG-only quality option to PNG removal. Six control commands and twelve marked-file commands failed before remover inference. Every command, error and completed encoding result is retained.
After freezing that run, a separate continuation plan registered the same eighteen PNG files by hash. It omitted the JPEG-only option and performed removal without re-encoding, replacing a payload or repeating a JPEG trial. All eighteen operations completed. The final removed-state counts combine the original JPEG results and that PNG continuation.
Image cost
RGB PSNR against the marked input ranged from 42.58 to 51.68 dB after pixel removal. The files changed, including unmarked controls. Dimensions, alpha presence, file and pixel hashes, mean squared error, PSNR and both decoder outputs remain in the record. This numeric range supplies no perceptual-quality guarantee.
Zero exact recovery establishes what these decoders returned for these outputs. It does not establish that another detector would find nothing, that arbitrary watermarks have gone or that the method applies to a provider's production system.
Runtime checks
The pre-inference advisory check found affected Rustix and TLS dependencies in the upstream lock. The experiment updated compatible versions before inference and retained both locks and both scans. The patched 316-package Rust/Python query retained only an informational maintenance notice for paste. Model hashes, upstream source revision, binary identities and MIT terms are recorded separately.
The 13 November research record links the complete original and continuation registrations, retained files and verifier. The verifier checks saved bytes, pixels, classifications and relationships without a new inference run.