The storage system doing this wouldn’t use that part of the hash, it would do it itself so no issues? (Hash chunks, instead of feeding everything in linearly)
Otherwise the hash isn’t going to be even remotely safe for most inputs?
You could consider building this storage system on top of BLAKE3's tree model. Namely you store blocks as small Merkle tree. And an image is basically a collection of blocks that has a different 'hat' on top of it. Unfortunately, BLAKE3 makes this hard, because the same block will end up having a different Merkle tree node depending on the offset at which it's stored.
Either modify the algorithm to your needs, and rename it.
Or just use something thats already suitable off-the-shelf. Plenty of merkle-trees out there.
Consider a crypto currency pow that did that without the chunk counter. It'd be trivially exploitably by precalculating all the tree but the chunk that changed per nonce.
Most OSes use 4K block sizes. To get the best dedup you have to hash every 4K block and lookup each one individually in a dedup table. Two VM images could both contain an identical 4GB file, but every 4K block of that file could be stored at different offsets in the VM images. A tree hash wouldn't let you dedup anything but identical sections stored at identical offsets, whereas a dedup table and 4K blocks allows you to dedup the entire file.
Trying to use Blake3's tree structure directly to dedupe is a misunderstanding of the problem you're trying to solve. Removing the counter would not let you use Blake3 alone for this purpose.
Though this feature is well intended, it makes this hash function hard to use for a storage system where you try to do aggressive data deduplication.
Furthermore, on platforms that provide native instructions for SHA hashing, BLAKE3 isn’t necessarily faster, and certainly more power hungry.