Hardware Requirements

Operating system

• Windows 10 (1903 or later) or Windows 11, 64-bit

Storage

• ~2.5 GB free disk space (the app unpacks to about 2.1 GB; the download itself is roughly 1.35 GB zipped)

GPU

• Any DirectX 12–capable GPU works: NVIDIA, AMD, or Intel, via DirectML. There's no vendor lock-in and no separate CUDA/ROCm install needed.
• No dedicated GPU? No problem. The app automatically falls back to CPU if no compatible GPU is found, and will also fall back mid-run if a GPU operation fails (e.g. a very low-VRAM card running out of memory on a large page) - it retries on CPU rather than crashing.
• CPU-only is noticeably slower. In our own testing, translating one manga page (detection + OCR + inpainting) took ~9s on GPU vs. ~79s on CPU (~9x slower) on the same machine. Fully usable, just not fast.

RAM

• The backend process itself measured ~700 MB at idle (models loaded, warmed up) and ~1–1.5 GB while actively processing a page in our testing. Budget for 8 GB system RAM as a comfortable minimum once you account for the reader UI, your browser/other apps, and OS overhead.

VRAM (if using a GPU)

• With detection, OCR, and inpainting all loaded at once, we observed GPU memory usage reaching up to ~11 GB on a 12 GB card during a full translation run. That number includes whatever else was using the GPU on the test machine at the time, so treat it as a rough upper bound rather than an isolated measurement - actual usage will vary by page size and what else is running.
• If your GPU is short on VRAM, the app will still work - it degrades to CPU for whichever step runs out of memory instead of failing outright.

Benchmarks

Test machine: NVIDIA GeForce RTX 5070, Intel Core i9-9900K (8C/16T @ 3.6GHz), 64GB RAM.

Full page translation (detect + OCR + translate + inpaint)

Backend	Time (test2.webp, 51 text blocks)
GPU (DirectML)	8.66s
CPU (no GPU)	78.89s
Speedup	~9.1x

Detection only (DBNet)

Image	Size	Text blocks	PyTorch/CPU	ONNX/DirectML	Speedup
test.webp	1280×1791	20	0.851s	0.211s	4.03x
test2.webp	1280×1943	51	0.878s	0.220s	4.00x
test3.webp	1280×1280	11	1.040s	0.189s	5.50x

Per-stage breakdown (warm, higher detect resolution)

Same test machine, but with detect_size raised to 2560 (larger working resolution than the table above) and measured warm - i.e. after every model is already loaded, isolating steady-state inference time from one-time model-load cost.

Stage	CPU	GPU (DirectML)	Slowdown
Detection	5.10s	2.46s	2.1x
OCR	76.03s	1.43s	53x
Inpainting (mask refine + Lama)	23.18s	6.54s	3.5x
Total	107.30s	15.92s	6.7x

OCR is the dominant cost on CPU by far - 53x slower than GPU, and 71% of the total CPU time. Detection and inpainting are comparatively minor (2.1x/3.5x), so if you're stuck on CPU, OCR is where the wait actually comes from, not detection or cleanup.