Add up the component table in the dev.to guide for a library with no video and it comes to 1.15x to 1.25x: originals at 100 percent of what you handed over, because nothing is recompressed [4], thumbnails and previews at 15 to 25 percent [5], and a Postgres volume in the hundreds of megabytes at this scale [7] [1]. The stated floor is 1.35x [1]. That difference is headroom rather than a fifth directory, which matters when your own measurements come in under the band.
The top of the band is video, and it lives in one directory. `encoded-video/` costs 50 to 100 percent of your video bytes again, and only for clips the browser cannot play natively [6]. Run the guide's own figures forward. A 12 MP HEIC frame lands near 2 MB [11], so 20,000 stills is about 40 GB [2], while the same 20,000 item phone library is put at 60 to 120 GB of originals [10]. That leaves 20 to 80 GB of video [3], which at 400 MB for a 4K clip [11] is between 50 and 200 clips [4]. Transcode those and you add 10 to 80 GB [5]. So the multiplier is a claim about codec mix, and for it to transfer your clips have to be the same kind of unplayable as the author's.
The 200 GB provision for a 90 GB Takeout [8] survives checking, and it is the 80 percent rule doing the work rather than the multiplier. Worst case footprint is 1.6 x 90, or 144 GB [1]. Holding that under 80 percent needs a volume of at least 180 GB, and 200 GB lands at 72 percent [6]. The stronger instruction in the same guide is to run `du -sh` on each subdirectory after the first 1,000 assets and derive your own multiplier [9]. That is the right shape for the problem: sample your library instead of trusting anyone's band, this one included.
Which leaves the sequencing. The three settings that force a full re-run are the storage template, the machine learning model and the transcoding policy, and all three are set in the first hour [3]. The measurement that would inform them arrives 1,000 assets later [9]. Import a thousand files, read the directories, then commit, and treat the first 72 hours as unattended processing rather than evidence of a broken install [17].
For the RAW case the arithmetic closes the question before policy comes up. 20,000 files at 25 MB is 500 GB, and 1.35x to 1.6x of that is 675 to 800 GB [14] [7]. The guide points that profile at an external library in read-only mode so Immich never becomes the only copy of the originals [14], which is right for a reason beyond disk: a read-only external library has no storage template to get wrong.
In my context I would take the low-power path, facial recognition on and the smart search model off, as the Pi 5 and N100 advice has it [15]. The honest cost of that choice is that the model is one of the settings which rewrites finished work when you change it [16], so adding smart search in month three means paying the queue you avoided in week one.
