AOT Inpainting for Manga

A SafeTensors conversion of the AOT inpainting generator released by zyddnys/manga-image-translator. It reconstructs masked regions after text removal from manga and comic images.

Model details

• Architecture: AOTGenerator
• Input: four channels (mask + RGB image)
• Output: three RGB channels
• Base channels: 32
• AOT blocks: 10
• Dilation rates: [2, 4, 8, 16]
• Tensor count: 168
• Format: Candle-compatible SafeTensors

Files and provenance

• model.safetensors: converted floating-point checkpoint with the original tensor names
• config.json: architecture and preprocessing metadata
• Source checkpoint: inpainting.ckpt
• Source SHA-256: 878d541c68648969bc1b042a6e997f3a58e49b6c07c5636ad55130736977149f

Intended use and limitations

The model is intended to fill binary-masked regions in illustrated images, especially regions left after comic text removal. It is not a general-purpose photo restoration model. Large images may need resizing or tiled inference; the supplied configuration uses a default maximum side of 1024 pixels. No evaluation metrics or training dataset artifacts are included in this conversion.

License

MIT, as declared for this converted model repository. Review the upstream project and source checkpoint terms when redistributing derived artifacts.