Writing strings to TDMS that are a byte stream can be tricky. I have a Read/Write Cluster to TDMS and it flattens it to a string. But as you noticed there needs to be special consideration for null. Well it also turns out that on different platforms there are other characters that need special attention too. You can absolutely do extra work on your images. But I fear that you won't get the performance you want. TDMS is crazy fast and all, but it sounds like a bonkers amount of data real fast, and I don't know if it will be able to keep up, especially if there is special code needed to convert a string ready to be written to TDMS so the null (and any other characters you want) get written correctly.
Writing an array of bytes also sounds crazy for this application. The number of samples on the channels might approach NI's limitations for the data type, and then keeping track of where the offsets are for files would need to be a separate channel. I've done stuff like this for logging raw CAN frames to TDMS files but that seems like way less data. Honestly you might be better served writing to binary files, then having the TDMS file just keep an index of the file names or locations for replay. Not sure how I'd go about solving this.