A frame grab is a compressed, motion-affected still

A single frame pulled from a video looks like a photo once you have extracted it, but it was not made like one, and a scoring model can tell the difference even when a person cannot.

Three things a frame carries that a photo does not

Video is compressed far more aggressively than a still photo, because a video file has to store many frames a second rather than one, and most video codecs only encode full detail on certain frames, filling in the rest with predicted differences from their neighbours. A frame extracted from the middle of that sequence can carry compression artefacts well beyond what an equivalent still JPEG at the same visual resolution would show, simply because the codec was optimised for motion, not for a single frame ever being examined closely on its own.

Many phone and camera sensors read out an image line by line rather than all at once, which is invisible when nothing is moving but produces a subtle skew or wobble - rolling shutter - when either the subject or the camera moves during that readout. A still photo taken with a normal shutter does not have this problem in the same way; a video frame, especially of a moving subject, often does.

Motion blur is also more common in extracted frames than in still photos, because video is typically shot at a fast enough frame rate that each individual frame gets comparatively little exposure time, and any movement during that short window smears across the frame the same way it would in a rushed handheld photo.

Why this reaches the score

All three of these are exactly the kind of low-level artefact a vision model's early layers respond to, whether or not anyone intended them to matter. Compression blocking reads as texture. Rolling-shutter skew reads as a kind of geometric distortion, however slight. Motion blur reads as reduced sharpness on whatever moved. None of these are properties of the subject, and all of them are properties of how the image was captured and encoded, which means a video frame and a still photo of the same subject, taken at the same moment, are not the same input to a model even if they look nearly identical to a person glancing at both.

What follows

Comparing a score from a still photo against a score from a video frame is not a controlled comparison, because the capture method itself changed, on top of anything else. If you want a result to mean something across repeated attempts, keep the capture method the same each time - running a comparison that actually holds variables steady covers this discipline in full, and capture method belongs on the list of things to fix, not vary. This overlaps with, but is distinct from, ordinary compression differences between file formats - how JPEG compression itself leaves marks a model can see is a separate, format-level question, and video compression is a more aggressive version of the same underlying issue rather than the same thing.

A specific tool's accepted formats and whether it takes video at all is a product question, not a mechanism one - Rate Cock is one example of a tool built around still photo uploads, which is the more common design across the category. Penis Rater's tool notes look at how different scoring tools vary in what they accept, and Measure My Cock's method pages note that a still, properly framed photo is also the better starting point for a physical measurement, for exactly the reasons above. A human reviewer working from a video can usually just watch it and mentally set aside a bad frame - Rate Penis's judging process is not locked to a single extracted still the way a scoring pipeline is.