Novel Mill is a translator’s aid, not a finished translation. It is designed to run alongside a manual English pass—or before you write one.

Novel Mill processes a Japanese web novel or light novel through a local AI model, one scene at a time, displaying the English translation next to the source text so the two can be easily compared.

Why Context Matters

A translation is not a lookup table. The same line can read as polite, cold, joking, or evasive depending on who is speaking and what just happened. Two careful translators will rarely land on the same sentence because context is subjective.

An AI pass offers another reading of that context. While it is often wrong, it can be useful in unexpected ways—it might flatten a joke you kept, or preserve a subtle implication you smoothed out. Displaying the original and AI translation side by side keeps choices visible rather than hidden.

                  [ Japanese EPUB / Scans ]
                              │
                              ▼
                       [ Novel Mill ]
           (Scene Splitter + Lock File Injection)
                              │
                              ▼
                [ Local LM Studio / Tesseract ]
                              │
                              ▼
                [ out/*.en.txt Side-by-Side Draft ]
                              │
                              ▼
                  [ Your Manual Edit Pass ]

Who Is This For?

• Built for: Fan translators, localizers, and language learners who want an offline, side-by-side draft to challenge their choices and accelerate manual editing.
• Not for: Anyone looking for a one-click automated publisher, a web novel scraper, or a DRM remover. You bring the book; Novel Mill helps you tackle the translation.

How It Works in Practice

1. Side-by-Side Translation Drafting

Instead of replacing your translation, Novel Mill gives you a draft to argue with:

Japanese Source:

「べ、別にあんたのために作ったわけじゃないんだからね！」

Novel Mill Output (out/01.en.txt):

“I-it’s not like I made this for you or anything!”

Your Final Manual Pass:

“Don’t get the wrong idea—I didn’t cook this for your sake.”

2. Character & Tone Consistency (Lock Files)

To stop the local model from changing character names or misinterpreting speech registers halfway through a chapter, optional JSON lock files let you lock down names, speech marks, and overall tone across every scene.

JSON

{
  "names": {
    "佐藤": "Sato",
    "美咲": "Misaki"
  },
  "speech_marks": {
    "「": "“",
    "」": "”"
  },
  "tone": "Casual dialogue, maintain light novel honorifics (-san, -kun), avoid over-localizing idioms."
}

Prerequisites & Dependencies

Neither dependency is bundled with Novel Mill:

• Text EPUBs: Require LM Studio only.
• Scanned Books: Require LM Studio and Tesseract.

1. LM Studio (Local Model Server)

LM Studio runs models locally on your hardware and exposes an OpenAI-compatible API. Novel Mill talks exclusively to that local API. Nothing in your book folder is ever sent to a cloud host.

• Default / Recommended Model:gemma-4-12b-it
  • Why it works so well: Built on Google’s 12B encoder-free unified architecture, it strikes an exceptional balance between VRAM footprint and translation quality. It excels at preserving modern conversational Japanese registers, subtext, and natural light-novel dialogue flow without over-flattening character personality.
  • System Requirements: Fits comfortably in ~10GB to 14GB of VRAM (4-bit/5-bit GGUF quantization).
• Installation:
  • Windows: Download the installer from lmstudio.ai or run:
    Bashwinget install --id ElementLabs.LMStudio -e
  • Linux: Download the AppImage from lmstudio.ai, mark it executable, and run it.
• Configuration:
  1. Open LM Studio, search for gemma-4-12b-it, and download a GGUF quant (Q4_K_M or Q5_K_M recommended).
  2. Load the model with Context Size: 8192 and Parallel Slots: 1 (4096 is too small once system lock files are loaded).
  3. In the Developer tab, start the local server (default: [http://127.0.0.1:1234](http://127.0.0.1:1234)).
  4. Ensure the model ID configured in Novel Mill matches the exact ID string shown in LM Studio.

2. Tesseract (OCR Engine)

Tesseract extracts Japanese text from scanned page images. Skip this step if your EPUB already contains text. Novel Mill requires two language files: jpn (horizontal text) and jpn_vert (vertical text). An English-only setup will not work.

• Ubuntu / Debian:
  Bashsudo apt install tesseract-ocr tesseract-ocr-jpn tesseract-ocr-jpn-vert
• Windows:
  1. Install Tesseract:
     Bashwinget install --id tesseract-ocr.tesseract -e
  2. Confirm jpn.traineddata and jpn_vert.traineddata exist in your tessdata folder (usually C:\Program Files\Tesseract-OCR\tessdata).
  3. If missing, copy both files from the official tessdata repository into tessdata.
  4. Verify setup by running tesseract --list-langs. Output must list both jpn and jpn_vert.

Model Benchmarks & Alternatives

While gemma-4-12b-it is the default recommendation, local LLM performance depends on your GPU VRAM budget and preferred translation style. Here is how alternative models compare for Japanese web novel translation:

Model	Size / VRAM Needed	Pros	Cons
Gemma 4 12B Instruct (Default)	12B (~10–14 GB VRAM)	• Superior handling of Japanese dialogue registers • Fast generation speed • Low hardware requirements	• Can occasionally hallucinate idioms if prompt instructions are too sparse
Qwen 2.5 14B / Qwen 3 14B	14B (~12–16 GB VRAM)	• Extremely strict adherence to JSON lock files • Outstanding literal accuracy for complex kanji compounds	• Tends toward rigid, formal English prose that requires heavier manual editing
Command R (35B)	35B (~22–26 GB VRAM)	• Exceptional contextual reasoning and flow • Designed specifically for multilingual tasks	• Requires high-end GPU hardware (24GB+ VRAM) • Slower scene generation times
Gemma 2 9B / Qwen 2.5 7B	7B–9B (~6–8 GB VRAM)	• Runs smoothly on budget GPUs or laptops • Lightweight VRAM footprint	• Struggles with context retention when using long lock files

Choosing the Right Model for Your Workflow:

• For Natural Prose & Character Voice: Stick with gemma-4-12b-it. It produces dialogue that feels closest to human localization.
• For Complex Worldbuilding & Heavy Lock Files:Qwen 2.5 14B (or newer Qwen 14B variants) follows rigid structural constraints better than almost any other model in its size class.
• For High-End Rig Users (24GB VRAM+):Command R (35B) provides rich, nuanced translations for dense prose, though generation speeds per scene will be noticeably slower.

Project Info

• Repository:github.com/PatchScratch/novel-mill
• Version on main: 1.8.0 (6 Oct 2026)
• License: MIT

Novel Mill is not a downloader, not a DRM remover, and not a substitute for reading Japanese.

.