Engine

The Engine tab in Settings picks which AI handles translation and how text is read from the screen. Mirrors the in-app Settings → Engine section. For the catalog of available models and their licenses, see Models.

Pipeline mode

Three pipelines, picked per session:

Text Reading — OCR → translate. Lightest, fastest, works on any model.
Text Reading + — OCR plus a structural pass that groups dialogue lines and merges fragmented text before translation. Balanced default, recommended for most games.
Image Recognition — the vision model reads the game screen directly, no OCR step. Best for stylized fonts and decorative UI. Requires a vision-capable LLM.

See Translation for the conceptual differences between modes.

Engine kind & model picker

Engine kind decides what runs translation:

Standard (Local LLM) — a large language model running on your GPU. Rich, context-aware translation, supports image recognition.
Lightweight (NMT) — a small CPU model for specific language pairs. Minimal VRAM, faster cold start, no image mode.

Tap Change to open the model picker. The picker shows download size, VRAM requirement, license, and whether the model is currently in use or already installed. Multiple models can coexist on disk; switching between downloaded models is instant.

Engine mode sits next to the kind selector — keep it on Default. Safe Mode forces a conservative profile (ngl=0 / ctx=2048 / no cache / no flash-attention) for when GPU init fails or VRAM is squeezed by another app.

OCR engine & speed

For pipelines that use OCR (Text Reading and Text Reading +), pick the OCR engine:

Windows OCR — built into Windows 10 / 11. Default. No extra download.
OneOCR — engine bundled with the Windows 11 Snipping Tool. Alternative for scenes where Windows OCR struggles.

OCR Speed trades accuracy for speed by downscaling the captured image before OCR. Auto picks a reasonable default; Normal never downscales; Fast / Faster cap the long edge at 800 / 600 px.