Also, could it be fake directory entries, like the ones used by LFN? Do simple 8+3 names for images make any difference? Parsing code should skip LFN's extensions, or collect all parts to construct the long name (Unicode will still pose a problem).