0.4.3

• Pangram Pangram's faces are drawn from the web fonts the foundry's site serves, by its permission: the crop's words in each face, counted by the second stage; first place on the foundry's own lines 27% → 35%.
• Training: the view pairs are trained at a share's memory (a bank of the batch), and the twelfth step is recorded.

Full Changelog: v0.4.2...v0.4.3

0.4.2

• The cover line "Aa" is a reference row of every face, in every catalog built from font files and in My fonts: Pangram Pangram's own product pictures, the design in the first ten for 63% of whole pictures and 62% of their largest lines, from 33% and 47%; 56 of 58 designs found from some picture, from 39.
• Lines are told apart however tight the leading: a crop of several lines no longer chains them into one.
• DaFont forum crops 50% top five from 48%; the other reads hold.
• Training checkpoints are judged on the photographs as the page reads them, raw and cleaned; --clean and --arc are training options.
• The catalog builder skips a face that hangs the renderer instead of stopping.

Full Changelog: v0.4.1...v0.4.2

0.4.1

• Every named weight and style of a variable font is a catalog row: PP Kyoto's Regular to Heavy and italics, PP Monument's widths and weights; Pangram Pangram from 351 rows to 4,580.
• A photograph is read as it stands and cleaned, its ink found by colour and drawn black on white; WhatFontIs-Bench final 49% top five from 43%, 26% first from 21%; screenshots read as before.
• A crop of several lines is read line by line, each family at its best line; a line set steeply on a slant is read turned level as well.
• Pangram Pangram's own product pictures: the design in the first ten for 33% of whole pictures and 47% of their largest lines, from 11% and 17%.
• The same-text stage on the cleaned crop: 72% top five and 46% first on 400 final photographs, from 66% and 44%.
• Pangram Pangram's newer PP Kyoto and PP Mori builds.

Full Changelog: v0.4.0...v0.4.1

0.4.0

• A line reader, 156 KB, reads the crop's words, and the page sets the first 30 matches in them: on 400 final photographs first place 44% from 26%, top five 66% from 50%; createReader, rerank, matcher.embed in the package.
• Every result previewed in the words read from the crop; editing them re-ranks.
• The ninth-step model: rendered text 93% top five and 73% first from 92% and 72%; trained on the DaFont forum's 8,630 identified pictures, 95,798 web screen lines and 8,718 families.
• Pangram Pangram catalog, 201 families, by the foundry's permission; Paper Mono and the manual fonts in training.
• Photographs it never trained on: 43% top five from 41% on 8,955 images; held-out open families 74%, Pangram Pangram's held-out designs 90%.
• The forum read: 5,546 held-out "what font is this?" threads, 17.5% top five (bench/forum.json); the forum's development threads as training views.
• Open families split by design: a product page's cuts and its variable twin on one side of the held-out line.
• train.style_bench repin: pins move to a changed preparation module only with a byte-for-byte proof, recorded in repins.

Full Changelog: v0.3.0...v0.4.0

0.3.0

• New model, trained for photographs: views rendered on 383 CC0 surface textures with the benchmark's own camera, light and ink ranges, and on 18,506 lines cut from the top websites' pages.
• Photographs it never trained on: 41% top five on 8,955 images, 597 of the benchmark's 600 fonts indexed; forum requests 46% from 44%; held-out open families 74% from 72%; rendered first place 72% from 73%.
• Twin credit in every read: a clone in another catalog counts as the same design; the page folds a family's copies across catalogs.
• WhatFontIs catalog: the 226 benchmark fonts no other catalog held, from WhatFontIs's own renders; 69 Adobe Fonts titles matched by their store words.
• Web screens read: 28% top five on held-out sites, 41% on web fonts, from 13% and 20%; sites split by hash.
• Six-bit catalog rows read by every reader.
• Training within 32 GB: the process tree peaked at 20 GB.

Full Changelog: v0.2.0...v0.3.0

0.2.0

• New model: 64-number embeddings, 6-bit weights, 4-bit catalog rows; 335 KB gzipped, 158 scripts.
• Trained on 8,517 open families beside Google Fonts, on display effects and on photographs.
• Right family in the top five 92%, first 73%, on every Google Fonts family; other scripts 84%.
• Seven catalogs: Google Fonts, DaFont, Adobe Fonts, GitHub, Debian, Font Library, Other; 102,517 distinct families.
• Search every catalog at once; sources under 100 families fold into Other.
• A stored preview for every match outside Google Fonts, showing its family's name; DaFont's own previews.
• Address a font sample by its font and text; share links carry the model.
• Accuracy on photographs, forum requests, web screenshots and Wikimedia Commons photographs, on the page and in bench/.
• Font finders compared on owner, sources, runtime, openness and cost; Lens added.
• Pages pinned to their own stylesheets and modules.
• Model and catalog numbers written short: the model file 9% smaller.

Full Changelog: v0.1.0...v0.2.0