Skip to content

Page enhance (upscaling)

This page covers the admin side of server upscaling. For what Enhance does and how readers use it, see Page enhance.

Nothing runs until both steps are done.

  1. Open Settings, Library, Server upscaling and switch on Enable server upscaling.
  2. Pick the models readers may use. Models are not bundled. Each downloads from a Maki GitHub release when you tick it and save, or from its Download button, and Maki checks its SHA-256 before using it. Some models are non-commercial only, and the licences are listed in THIRD_PARTY_NOTICES.md in the Maki repository.

Readers then turn Enhance on under Enhance in the reader’s settings panel, where the Method list shows Maki 2x, Real-ESRGAN and SPAN-S under “On the server”. They still choose Off, Auto or Always for themselves, so enabling the feature changes nobody’s reading.

Upscaling runs on the server CPU. Maki ships the CPU build of ONNX Runtime and has no GPU option. A page takes several seconds of CPU, and a few times longer on a NAS.

  • Threads sets how many CPU threads one page uses, up to the number of cores (the default is half the cores, at most 4). A page is split into small tiles, and each thread works on one tile at a time.
  • JPEG XL for colour pages, on by default, saves upscaled colour pages as JPEG XL, so dark gradients stay smooth on an OLED screen and files are about 14% smaller. A browser that cannot show JPEG XL, such as Edge, gets WebP for that page instead. Grey pages are always WebP.
  • Cache size (MB) caps the disk cache of upscaled pages, 1 GB by default. The oldest pages go first. Clear cache empties it at any time.
  • Test speed times a built-in sample page and reports seconds per page, which is how long a reader waits for the server method.

Each model shows a speed and quality rating, and Maki 2x is marked Recommended.

Readers can skip chapters by file quality under Don’t enhance in their reader settings. Skipped chapters never reach the queue, so that work stays off the server. A reader who turns on Whole chapter queues every page of the chapter they have open and then the whole next chapter, so expect more work per reader with it on.

Prepare chapters ahead, off by default, lets the server upscale while nobody is reading. It takes the next chapter of each series a reader has set to Reading and read in the last 14 days, up to 8 series per reader, so they open sharp even on a phone. Only series whose reader settings have Enhance on Always with a server model count.

  • Read within (days), Series per reader, Chapters per series (1 to 3) and Space (MB) (4 GB by default, apart from the cache size) are adjustable.
  • Prepared pages stay until the chapter is read.
  • Reading always comes first: a prepared page in progress stops the moment someone opens a page.
  • Preparing runs on half the upscaling threads, so it takes about half the CPU and does not slow the page someone is reading.

Anyone who can edit a series can override this under Automation in its actions menu: follow Settings, never prepare it, or keep 1 to 20 chapters ready, for a series read in long sittings. A series with its own number is prepared however long ago it was last read and does not count toward the series limit.

Pages over 16 megapixels are never upscaled on the server, since one can take around 2 GB of memory. The reader uses a browser method for them.

Pages upscaled with Maki 2x before v0.37.0 are made again the next time they are read, because of a fix for faint boxes on dark panels and a higher quality for colour pages.

While the server works on a page, the reader shows a browser method instead, chosen under While the server works in the reader settings. Browser methods run on the reader’s device and need WebGL2. Without it, the page is shown plain.