Magent: Measured what the encoder quality setting costs, and why it is not the thing to change

2026-09-26T13:38:57Z

Measured what the encoder quality setting costs, and why it is not the thing to change

New page

'''Video quality and storage''' records a measurement taken in September 2026, and a decision not to act on it.

Every release is published as [[Cryptograss:Video encoding pipeline|four renditions]], and hunter's disk reached 89% full. The obvious lever is the encoder's quality setting. This page measures what that lever is worth, so the answer is on record rather than re-argued from intuition later.

'''The short version: leave it alone.''' All published video comes to about 2 GB. The backup mirror on hunter is 53 GB. Degrading the masters to save a few hundred megabytes would be the wrong trade.

== What was measured ==

The 1080p rendition — the largest of the four, and the one worth arguing about — encoded at CRF 32 (what ships today) through 40, on two real releases at opposite extremes:

* '''Busy''': [[Release:QmRpqcstuYx2yhZ5TjXdVHSUefwDGrPoe65sNTkuNUGkqz|Red Haired Boy — Görlitzer Park]], 28s, a park full of grass and leaves.
* '''Quiet''': [[Release:QmZSLe6SkfRB65DdyYeQFVyF1L6p7pFbyk1t4TCV65o4pB|Victory Theme at a Nashville Airbnb]], 29s, a still room.

{| class="wikitable"
! Setting !! Busy clip !! Quiet clip !! XPSNR (busy) !! Looks like
|-
| '''CRF 32''' (current) || 42.4 MB || 3.5 MB || 38.65 || reference
|-
| CRF 34 || 37.3 MB (−12%) || 2.9 MB (−17%) || 38.01 || no visible difference
|-
| '''CRF 36''' || 32.4 MB (−24%) || 2.3 MB (−34%) || 37.33 || no visible difference
|-
| CRF 38 || 27.9 MB (−34%) || 1.8 MB (−49%) || 36.71 || grass begins to soften
|-
| CRF 40 || 24.0 MB (−43%) || 1.5 MB (−58%) || 36.12 || visible smearing on foliage
|}

[[File:Video quality crf comparison.png|thumb|none|900px|Left to right: CRF 32, CRF 36, CRF 40. Grass and denim at twelve seconds in.]]

Quiet footage compresses far better than busy footage at the same setting — 58% saved against 43% at CRF 40 — because there is less real detail to keep. Encode time barely moves with CRF; that is the ''preset'', not the quality target.

== Why the answer is still "leave it" ==

{| class="wikitable"
! What !! Size
|-
| Every published video, all renditions || '''~2 GB'''
|-
| Backup mirror on hunter || 53 GB
|-
| Docker build cache (reclaimed by hand, September) || 14 GB
|-
| Unused Docker images || 25 GB
|}

A 24% saving on 2 GB is 500 MB. [https://github.com/cryptograss/maybelle-config/pull/134 Retention on the backup mirror] returns about 46 GB — ninety times as much, at no cost to anything anyone will ever watch.

The archive exists to be the durable copy. Spending picture quality to save half a gigabyte, while ten times that sits in build caches, would be solving the wrong problem with the only irreversible tool available.

== When to revisit ==

* The published archive reaches tens of gigabytes — likely once several hour-long sets go up. At that point CRF 36 looks close to free.
* Someone reports buffering on the 1080p rendition. The fix there is bitrate, and this table says what it costs.
* A source turns out to be badly served by CRF 32 — very dark footage, heavy grain — in which case the setting should vary by content rather than globally.

'''One number worth keeping''': 23 GB of source video has been uploaded to produce those ~2 GB of published renditions. The sources are deleted after publishing. If storage pressure ever does come from video, that ratio is the more interesting end of it.

== Caveat on the method ==

delivery-kid deletes the camera original once a release is pinned, so these encodes were made from the ''published'' 1080p renditions rather than from phone files. Absolute sizes therefore run lower than a fresh upload would produce, and the quality scores compare against the published copy rather than the camera's.

The shape of the curve — what each step of CRF saves — is unaffected by that, and it is the part the decision rests on. Anyone wanting absolute numbers should re-run it against a camera original, kept aside before publishing.

== See also ==

* [[Cryptograss:Video encoding pipeline]] — the four renditions and why AV1
* [https://github.com/cryptograss/maybelle-config/issues/131 maybelle-config#131] — the disk that prompted the question

Cryptograss:Video quality and storage - Revision history

Magent: Measured what the encoder quality setting costs, and why it is not the thing to change