micro-manga-translator models

The model files loaded by the micro-manga-translator web page, which detects, recognizes, translates (Japanese → English), erases and re-letters manga text entirely in the browser on WebGPU with onnxruntime-web.

folder	model	source	license
ppocr-manga/	PP-OCRv6_manga v0.3 text detector + recognizer (unmodified)	Kellenok/PP-OCRv6_manga	Apache-2.0
migan/	MI-GAN inpainting, ONNX pipeline export (unmodified)	Picsart-AI-Research/MI-GAN via anyisalin/migan-onnx	MIT
sugoi-v4/	Sugoi v4 ja→en, converted to fp16 ONNX (encoder + decoder)	entai2965/sugoi-v4-ja-en-ctranslate2	NTT terms of use: research use only, no commercial use

Each folder has its own LICENSE; the terms of that file apply to the files in that folder.

Sugoi v4 conversion

The ctranslate2 model.bin was parsed directly, the fairseq transformer (6+6 layers, d=1024) rebuilt in PyTorch, checked token-for-token against ctranslate2, and exported as two ONNX graphs:

• encoder.onnx: input_ids[B,S], attention_mask[B,S] → hidden[B,S,1024]
• decoder.onnx: tokens[B,T], hidden, attention_mask → log-probs of the last position [B,V] (no KV cache)
• tokenizer.json: sentencepiece unigram pieces/scores and model vocabularies

These files are derived data of NTT's model and are under the same terms: they may be used only for research purposes involving information analysis, and not commercially.