The problem lies in what is a "depiction":

(5) "visual depiction" includes undeveloped film and videotape, data stored on computer disk or by electronic means which is capable of conversion into a visual image, and data which is capable of conversion into a visual image that has been transmitted by any means, whether or not stored in a permanent format;

section 2256(5) of title 18

https://uscode.house.gov/view.xhtml?req=(title:18%20section:2256%20edition:prelim)%20OR%20(granuleid:USC-prelim-title18-section2256)&f=treesort&num=0&edition=prelim

via: section 1309 of the Consolidated Appropriations Act, 2022 (15 U.S.C. 6851).

https://uscode.house.gov/view.xhtml?req=(title:15%20section:6851%20edition:prelim)

via: the definitions section of the act

https://www.congress.gov/bill/118th-congress/senate-bill/4569/text#idE946FA7637914C2F88ACBEDF472397DB

The way it is written, even cropped, rotated, blurred, or in any other way processed files of that "depiction", even the values learned by a neural network (capable of conversion into a visual image), would fall under the "identical" part.

Since perceptual hashing does exist, there are open source libraries to run it, and even Beehaw runs an AI based image filter, the "reasonable effort" is arguably to use all those tools as the bare minimum. Even if they sometimes (or always) fail at removing all instances of a depiction.

But ultimately, deciding whether a service has applied all "reasonable efforts" to remove "identical copies" of a "depiction", will fall on the shoulders of a judge... and even starting to go there, can bankrupt most sites.