Skip to content
Manga Translator Wiki

English

简体中文
简体中文

English

简体中文
简体中文

Appearance

CLI, Batch, and Output ​

This guide covers the CLI, batching, and output fields in General settings and explicit overrides from the local local CLI. It does not cover API credentials, detection/OCR/translator/typesetting algorithms. Configuration and screenshots must be sanitized: never show real keys, tokens, usernames, private absolute paths, user images, or private prompts.

What these settings control ​

This page includes logging, error isolation, GPU/ONNX, retries, translation context, batch/pipeline scheduling, format/quality/overwrite, text/JSON/TXT sidecars, source-directory output, PSD/JSX, the custom API-parameters file, and model cleanup after translation. context_size is actually in the Translation tab; workflow labels belong to the translation workspace, so this page records only their batch constraints.

Change it in the desktop app ​

Open Settings and select the General tab. Numeric values use line edits, booleans use toggles, and format uses a combo box. The General layout comes from tab_custom_1 in settings_tab_layout.json. Changes update memory and are merged to disk; importing configuration or switching presets can rebuild rows. The Edit action beside use_custom_api_params opens a JSON file; it is not another core configuration value.

The formal local CLI entry point is manga_translator/args.py: python -m manga_translator local -i <sanitized input> [-o <sanitized output>]. -i/--input accepts multiple values; --config, -v/--verbose, --overwrite, --use-gpu, --disable-onnx-gpu, --format, --batch-size, and --attempts can override configuration, but an omitted value does not override it. The formal top-level subcommands are local, web, ws, and shared; this page discusses only local input/output overrides.

Parameters ​

For the mapping of UI names, storage keys, and default values of the parameters on this page, see the Settings Parameter Index.

Verbose Logging ​

The “Verbose Logging” toggle is on Settings → General. Qt UI always creates result/log_<timestamp>.txt at startup, even when this toggle is off. Enabling it raises console output to DEBUG level and creates a per-image debug folder containing cached intermediate input, detection/OCR, inpainting, and rendering artifacts. Results are unchanged, but the extra files can be numerous and consume substantial disk space; use it only for troubleshooting. Default: false.

Retry Attempts ​

“Retry Attempts” is an integer input. It sets how many times a translation request is retried after failure; -1 means unlimited. It applies only to the translation/API request layer, not to a general detector/OCR/render retry, and it is not the high-quality retry. Default: 3.

Ignore Errors ​

When the “Ignore Errors” toggle is enabled, a failed image is recorded and processing continues with the remaining images; it does not turn failure into success and does not swallow cancellation. Default: false.

Use GPU ​

The “Use GPU” toggle decides whether model loading and inference use the GPU. It requires matching drivers, Torch/CUDA, and VRAM; enabling it does not guarantee that every implementation uses the GPU. Default: true.

Disable ONNX GPU Acceleration ​

The “Disable ONNX GPU Acceleration” toggle forces only ONNX sessions onto the CPU and does not disable Torch CUDA. It can avoid provider conflicts at a speed cost. Default: false.

Context Pages ​

“Context Pages” is on the Translation tab and is an integer input. The translator builds context from recent non-empty history pages; 0 or a negative value disables it. See Context and Prompts for details.

Batch Size ​

“Batch Size” is an integer input that controls how many images are submitted per translation batch, the concurrent queue bound, and the memory peak; it is not the number of images running at once, and special modes may force 1. Default: 3.

Concurrent Batch Processing ​

When the “Concurrent Batch Processing” toggle is enabled, the detection+OCR, translation, inpainting, and typesetting stages run in parallel through queues. It is stage-level parallelism, not all images requesting the API simultaneously. TXT import, JSON-only, original/translation export, colorize/upscale/inpaint-only, and replacement translation force concurrency off.

flowchart LR
    subgraph Off["Concurrency off"]
        O1["Image 1: all stages"] --> O2["Image 2: all stages"]
    end
    subgraph On["Concurrency on"]
        A["Image A: detection + OCR"] --> B["Image A: translation"]
        A --> C["Image B: detection + OCR at the same time"]
        B --> D["Image A: inpainting + typesetting"]
    end

This does not mean every image sends API requests simultaneously; the queue and batch size provide backpressure, and special workflows disable the mode. Default: false.

Output Format ​

“Output Format” is a combo box on Settings → General that decides the extension used when saving images.

  • Not Specified: preserve the original extension.
  • png: PNG.
  • jpg/jpeg/jfif: JPEG (requires RGB conversion).
  • webp: WebP, supports quality.
  • avif: AVIF, requires Pillow codec support.
  • bmp: BMP (requires RGB conversion).
  • tiff/tif: TIFF.
  • heic/heif: HEIF, requires codec support.

Default: Not Specified.

Overwrite Existing Files ​

When the “Overwrite Existing Files” toggle is off, existing image output is skipped and TXT/JSON workflows check their corresponding files; results contain skipped items. Default: false.

Skip Images Without Text ​

When the “Skip Images Without Text” toggle is enabled, images with no translatable text after detection/OCR are skipped as a normal branch, not as an “Ignore Errors” exception. Default: false.

Editable Image ​

When the “Editable Image” toggle is enabled, JSON sidecar data containing regions, original/translated text, dimensions, and rendering fields is exported; combined with “Export Original Text” it also exports the original TXT. Default: true.

Image Save Quality ​

“Image Save Quality” is an integer input applied to image, inpainting, and editor saves. Higher values usually make files larger; exact semantics depend on the encoder. Default: 100.

Save to Source Directory ​

When the “Save to Source Directory” toggle is enabled, results are written beside the source under manga_translator_work/result; otherwise the output directory is used and relative structure is preserved when possible. The directory must be writable. Default: false.

Export Editable PSD ​

When the “Export Editable PSD” toggle is enabled, the final stage creates Photoshop PSD/JSX; running the script requires Photoshop, and this is not an ordinary image format. Default: false.

Generate PSD Script Only ​

When the “Generate PSD Script Only” toggle is enabled, only the JSX script is generated and Photoshop is not run; the script can contain paths and must not be shared directly. Default: false.

Use Custom API Params ​

When the “Use Custom API Params” toggle is enabled, config/custom_api_params.json is read to attach extra request parameters; the adjacent Edit button opens that JSON file. See Custom Request Parameters for details.

Unload Models After Translation ​

When the “Unload Models After Translation” toggle is enabled, the desktop unloads models after a task completes, reducing resident memory at the cost of the next task’s loading time. Default: false.

How the settings take effect ​

UI/configuration file data goes through AppSettings/ConfigService, memory configuration, core Config, MangaTranslator, and export consumers. local in mode/local.py applies only explicitly supplied CLI values to cli; desktop app_logic.py additionally builds save_info with output directory, format, overwrite, and source-directory flags.

flowchart TD
    U["General or explicit local CLI parameter"] --> C["AppSettings / ConfigService"] --> M["Core Config"]
    M --> P["Detection + OCR"] --> T["Translation: context_size / attempts"] --> R["Inpainting + typesetting"] --> E["Image, JSON/TXT, PSD/JSX"]
    B{"batch_concurrent?"} -->|no| S["Normal batch by batch_size"] --> P
    B -->|yes and normal mode| Q["Four-stage queues and backpressure"] --> P

batch_size is the translation batch and concurrent translation queue bound; batch_concurrent is stage-pipeline parallelism, not an API concurrency count. TXT import, JSON-only, original/translation export, colorize/upscale/inpaint-only, and replacement translation disable it to preserve ordering and per-image file writes. context_size builds messages from recent non-empty pages; attempts, HQ quality retry, and API candidate rotation are separate layers; cancellation is not an ignorable error.

Interactions and caveats ​

GPU backends require compatible drivers, Torch/CUDA, and ONNX providers. Larger batches, concurrency, more context, and high-quality output increase resource, token, or disk pressure; for OOM, lower batch size or disable concurrency before increasing retries. Disabled overwrite skips images/TXT/JSON. Template and JSON-only workflows require matching work-directory files. PSD execution requires Photoshop; JSX/JSON/TXT can expose paths and text, so sanitize before sharing. Format encoding depends on Pillow and platform codecs.