LaMa Manga Inpainting

A SafeTensors conversion of an anime- and manga-finetuned Big-LaMa model. It fills masked image regions and is intended for removing text or other localized content from illustrated images.

Model details

• Architecture: Big-LaMa FFCResNetGenerator
• Input: four channels (RGB image + binary mask)
• Output: three RGB channels
• Base features: 64
• Downsampling stages: 3
• FFC residual blocks: 18
• Maximum features: 1024
• Format: SafeTensors with the original FP32 tensor values

The published checkpoint was described as fine-tuned on approximately 300,000 manga- and anime-style images and as an improvement over the older lama_mpe model.

Files and provenance

• lama-manga.safetensors: 989 tensors converted from the original TorchScript checkpoint
• config.json: architecture and preprocessing metadata
• Source checkpoint: anime-manga-big-lama.pt
• Base architecture: advimman/lama

The conversion removes only the TorchScript wrapper prefix (model.generator.); tensor shapes, dtypes, and values are preserved.

Intended use and limitations

Use this model for masked inpainting of manga, anime, and comic artwork. Masks should identify pixels to replace. Results depend on mask quality and surrounding context, and the model can invent textures or line art. It should not be used where reconstructed pixels must be treated as factual source content. No evaluation metrics or training dataset artifacts are included here.

License

MIT, as declared by this model repository. Review the LaMa implementation and source checkpoint terms for downstream redistribution.