The container matters less than the re-encoding

A scoring model never sees "a HEIC file" or "a PNG." By the time the image reaches it, everything has been decoded into the same kind of pixel grid regardless of what format it arrived in. The format itself is not the variable that moves a score. What moves it is what happened to the pixels on the way there.

Decoding erases the format

HEIC, JPEG and PNG are containers with different rules for storing colour data, but once a decoder reads the file, the result is a grid of pixel values in memory. A model trained to expect that grid does not carry any record of which container the pixels came from. If a HEIC photo and a PNG of the exact same pixels were both fed to the same model, the scores would match, because from the model's point of view they are the same image.

Where the real difference comes from

The catch is "the exact same pixels" rarely holds in practice. JPEG is a lossy format: every time an image is saved as JPEG, its pixel values are recompressed, and recompression discards information the previous save may have already discarded once. Converting a photo from HEIC to JPEG, or re-saving a JPEG through an editor, re-encodes it, and each re-encoding pass introduces small, cumulative changes - blocking, softened edges, minor colour shifts. These are the same kind of artefacts covered in more depth in how JPEG compression marks show up to a vision model; this note is about the conversion step that produces them, not the artefacts themselves. PNG is lossless, so converting to PNG does not degrade pixels further, but converting from a lossy JPEG into PNG only preserves whatever damage the JPEG step already did. PNG does not undo it.

What this means practically

A single format choice does not predict a score change. A chain of conversions does, because each hop is another chance to re-encode. The photo most likely to score differently from itself is not "the JPEG one" - it is the one that has been saved, opened, edited and re-saved several times across different apps, regardless of which format it ends up in. If you are testing whether format affects your results, the fair comparison is one decode versus one decode, not one file that has passed through three tools against one that has passed through none.

What this is not about

This note does not cover metadata handling, which is a separate question about what travels inside the file rather than what happens to the pixels - privacy handling for uploads covers that side. It also is not the place to go deep on compression artefacts themselves, since that already has its own post. Rate Cock accepts the common phone formats without asking you to convert anything first, which sidesteps the conversion-chain problem by not requiring one. The physical-measurement side of file handling, where a consistent method matters more than a consistent container, is covered on Measure My Cock, and reading how a tool describes its own accepted formats is one of the smaller checks Penis Rater recommends before trusting a comparison across uploads. A human reviewer never sees a file format at all, only whatever image a submission tool renders on their screen, which is one of the ways a human judge's process differs structurally from a model's.