v1.70 — OCR tune-up: fewer misreads, faster rescans

Game Overlay Translator · published · build 26222478

Second OCR-quality update, built on measurements like the last ones - plus a fix for a long-standing model-folder bug that two players reported this week.

NEW: Upside-down misreads are gone. The angle check that could flip a normal horizontal line and turn a HUD "28" into "82" now runs only on steeply tilted text boxes - 24-29 misread lines per benchmark run became 0, slanted signs still read correctly, and OCR got 12-22% faster on dialogue and full screens (GPU).
NEW: Re-reading the same subtitle spot is much cheaper. When a subtitle changes in place, the app now skips the detection stage and only re-reads the characters - about 205 ms down to about 63 ms per re-scan (GPU) - with strict guards: if anything about the spot changed, it falls back to a full scan, so accuracy is untouched (zero regressions across 3,000+ benchmark images).
NEW: Fixed: your custom AI-model folder is now respected on every launch. Starting a translation right after a restart could launch Ollama without the configured folder, so models re-downloaded to drive C - up to 20 GB, as two players reported. Sorry. Now every path uses your folder, and if an Ollama you started yourself cannot see your models, the app tells you instead of silently re-downloading.

Honest notes: the angle change means genuinely upside-down (180-degree) text is no longer read - we found zero real cases across 924 game frames, but if your game has some, one config switch restores the old behavior. The detection skip only fires on same-spot swaps and re-checks; everywhere else nothing changes. And the benchmark verdict: players asked for the heavier "server" OCR models, so we tested them on 1,554 images - our current model beat them on quality, speed and language coverage, so it stays. Sometimes the honest answer is that the upgrade is not one.