Koharu: Rust-Native Manga Translator Disrupts Python's Stranglehold on Computer Vision

Pipeline Architecture

Koharu implements a five-stage computer vision pipeline entirely in Rust, avoiding the FFI overhead and GIL constraints typical of Python-based alternatives:

System Design

• Tauri Frontend: WebView-based GUI consuming <40MB RAM vs 300MB+ Electron baseline, with React/Vue bindings for the interface layer
• Zero-Copy Image Pipeline: Uses image crate and GPU texture sharing to avoid serializing arrays between detection → OCR → rendering stages
• Async Runtime: Tokio-based concurrency for parallel page processing and non-blocking I/O during translation API calls
• Modular Extractors: Plugin system supporting both local models (Sugoi NMT, Llama.cpp) and cloud APIs (DeepL, OpenAI) via WASM-compatible interfaces

Design Trade-offs

The Rust implementation sacrifices Python's vast ML ecosystem for distribution simplicity. While Python tools require Conda environments and CUDA toolkit alignment, Koharu ships as a single binary with statically-linked ONNX Runtime—trading model flexibility for operational reliability.

The Killer Innovation: Koharu is the first open-source manga translator to implement the entire CV/ML pipeline in systems-level code rather than orchestrating Python scripts. This eliminates the "Python environment tax" where users spend hours resolving torch and opencv-python version conflicts before translating a single page.

Technical Breakthroughs

• Rust-Based Typography Engine: Custom text layout algorithms handling vertical-rl writing modes, furigana ruby annotations, and font fallback chains for CJK characters—rendering directly to GPU textures without Cairo/Pango dependencies
• ONNX Runtime Integration: Native Rust bindings to ORT with DirectML/CUDA execution providers, achieving inference latency <50ms for text detection on consumer GPUs (RTX 3060) compared to 200ms+ in PyTorch-based implementations
• Incremental Processing: Implements sliding-window OCR with rayon parallelism, allowing real-time preview updates as users scroll through pages rather than batch-processing entire chapters
• Smart Inpainting Masks: Generates alpha masks that preserve screen tone (screentone) patterns during text removal, solving the "flat gray blob" problem common to粗暴 inpainting approaches
• Memory-Mapped Model Loading: Uses memmap2 for zero-overhead model weight loading, keeping RAM usage under 200MB even with 1.5GB parameter models (vs 2-3GB in Python equivalents)

Benchmarks vs. Python Competitors

Scalability Characteristics

• GPU Utilization: Efficient batching via ONNX Runtime allows saturating VRAM with multiple pages simultaneously, unlike GIL-limited Python implementations
• Concurrent Processing: Async architecture supports translating multiple manga volumes in parallel without spawning separate processes
• Limitations: Currently limited to ONNX-compatible models; cannot dynamically load PyTorch .pth checkpoints without conversion. Large language model integration requires GGUF format via llama.cpp bindings.

Competitive Landscape

Integration Points

• Model Ecosystem: Supports importing HuggingFace models via ONNX conversion, with built-in presets for manga-specific OCR (MangaOCR weights) and NMT (Sugoi Translator)
• Translation Backends: Pluggable architecture supporting local LLMs (via llama-cpp-rs), OpenAI-compatible APIs, and traditional NMT engines
• Export Formats: PSD layer preservation for professional typesetters, WebP/JPEGXL output for digital distribution, and JSON metadata for translation memory systems

Adoption Risks

The project is pre-1.0 and solo-maintained (mayocream). While the Rust implementation offers technical superiority, the Python ecosystem's network effects in ML mean Koharu relies on model conversions. The lack of CUDA-specific optimizations (depending on ONNX Runtime's generic kernels) may limit peak performance versus hand-tuned PyTorch inference.