Jiwami 1.0.3 - Huge UI, Manga, and Extraction Improvements
This update is mainly a complete overhaul on Jiwami's UI and design, but it also has major vocabulary extraction improvements, proper Settings customization, huge Manga accuracy and speed improvements, and a lot of smaller fixes and optimizations across Jiwami.
Complete UI and design overhaul
Complete overhaul to all the Screens.
Home:
Archive:
Settings:
Result Screen cards show all the definitions for the overview. Recent history is now synced with Archive, and Archive entries open in the full custom screen.
Result:
Archived:
Long file names now move to be readable instead of clipping.
Source rows, text sizes, spacing, icons, and metadata were cleaned up across the whole interface. Window sizes also got improvments and more flexability.
Settings presets and customization
You can now make proper Settings presets for Anime, Manga, Novels, or whatever workflow you want. Give them any name and color you want, select them later, update them, or make a new one from your current Settings. The selected one has its own animated color effect so you can tell what you are using immediately.
https://reddit.com/link/1vs2qyz/video/e2tyhc0y67kh1/player
If you change Settings without using a preset, you can just apply them normally. If you are editing a preset, you can update it. There is also a proper color picker now for presets, pitch accents, and target-word highlighting instead of being locked to a few colors.
More card control
Context furigana can now be shown on the front, back, both, or neither. The target word itself has a separate furigana setting, so you can show furigana for the whole context but hide it only for the mined word, or do the opposite.
Target-word highlighting can now be turned on or off, and you can choose its color. English definitions can now be turned off completely if you rather use the monolingual dictionaries only, or limited from 1 to 10 meanings if you do not want a card filled with too much English.
Vocabulary extraction improvement
Jiwami is now much better at handling dialectal, archaic, variant, and longer chain-heavy forms, with around 32% fewer measured errors.
Speed improvments are also a big part of this update. As i said before, extraction is done mainly with GPU. CPU also works, but it'll always be much slower. In my testing with a 667-line subtitle file, both the GPU and CPU gave the exact same final result: 656 Unique vocabulary.
My RTX 3060 Ti took 2 minutes, while the i5-12600 took 27 minutes, which is around 92% less time on the GPU for the same result. Of course this is one test on my own hardware, so do not take it as a universal number, and better hardware means even faster speeds.
Manga accuracy improvements
Manga text recognition now reaches 100% accuracy across 100 complex-layout Berserk pages. Berserk has very busy layouts, so this was not easy. Most of the text was already correct; the biggest improvements were punctuation and all the small details around it, bringing the number of mistakes down from 312 to 0 across these pages.
Manga speed improvements
The same 100 difficult Berserk pages used to take 28 minutes and 42 seconds. They now take 4 minutes and 48 seconds, which is around 83% less time or almost 6x faster while keeping 100% accuracy across the same pages.
Like always these numbers are from my own PC with these exact 100 pages, so they are not universal. Better hardware will be faster, but the same improvements are there for everyone.
For additional info, see the Jiwami changelog.