MangaShift External Resources

External models and code repositories for the MangaShift project.

🔗 Main Project: https://github.com/MattyMroz/MangaShift
🔗 Datasets: https://huggingface.co/datasets/MattyMroz/MangaShift

---

Naming Convention (models/ and .legacy/models/)

Every model in external/models/ and every curated legacy payload in external/.legacy/models/ follows one naming convention. Full reasoning: docs/archive/brainstorms/2026-04-27-naming-convention/.

The 11 rules

1. Casing: kebab-case, lowercase, ASCII only. No CamelCase, no underscores.
2. Format subfolders: each model gets pt/, onnx/, safetensors/, gguf/ per format inside external/models/<category>/<model-name>/, whenever that format exists.
3. Single-file model: <model-name>.<ext>, for example big-lama.pt, comic-text-detector.pt.
4. Multi-part model: one subfolder per variant, and the file name is the part name: <model-name>-<variant>/<part>.<ext> (hi-sam/onnx/hi-sam-b/encoder.onnx, magi-v3/onnx/vision-encoder.onnx).
5. Variant size: size goes in the suffix: <model-name>-<size>.<ext> (anime-text-l.pt, hi-sam-b.pth; sizes are n/s/m/l/x or b/l/h).
6. Quantization: suffix only when not FP32: <model-name>-<quant>.<ext>. Standard suffixes: fp16, bf16, int8, q4-k-m, q8-0. FP32 carries no suffix.
7. Upscaling exception: keep the scale prefix (4x-, 2x-). It is the community standard from OpenModelDB.
8. ComfyUI exception: models for the ComfyUI stack (flux, qwen) keep the diffusion_models/, text_encoders/, vae/ layout. The manifest classifies them as format comfy even when the files carry a .safetensors extension.
9. Upstream hashes: strip hashes from upstream names (sam_vit_b_01ec64.pth becomes sam-vit-b.pth). Versioning goes through bundle.yaml.
10. HF model.safetensors default: rename it locally to <model-name>.safetensors. Generic names are not used.
11. Source placeholder: every model has a src/ with a .gitkeep, even when the sources or export scripts are not stored yet.

Examples

external/models/
├── detection/
│   ├── anime-text/{pt,onnx}/anime-text-{n,s,m,l,x}.{pt,onnx}
│   ├── comic-text-detector/{pt,onnx}/comic-text-detector.{pt,onnx}
│   ├── hi-sam/
│   │   ├── pt/hi-sam-{b,l,h}.pth
│   │   └── onnx/hi-sam-{b,l,h}/{encoder,decoder,fg-decoder}.onnx
│   └── magi-v3/{pt,onnx}/<part>.{safetensors,onnx}
├── inpainting/
│   ├── big-lama/{pt,onnx}/big-lama.{pt,onnx}   # plus big-lama-fp16.onnx
│   └── flux2-klein/                            # ComfyUI exception
│       ├── diffusion_models/flux-2-klein-{4b,9b}.safetensors
│       ├── text_encoders/qwen-3-{4b,8b}.safetensors
│       ├── vae/flux2-vae.safetensors
│       └── src/.gitkeep
├── ocr/
│   ├── paddle-ocr/gguf/paddle-ocr-vl-{1.6,1.6-mmproj}.gguf
│   └── qianfan-ocr/gguf/qianfan-ocr-{bf16,q4-k-m,q8-0}.gguf
└── upscaling/
    ├── pt/4x-{anime-sharp-esrgan,...}.pth
    ├── onnx/4x-{anime-sharp-esrgan,...}.onnx
    └── src/.gitkeep

external/.legacy/models/
└── detection/
    ├── magi/{pt,src}/
    └── magi-v2/{pt,src}/

HF publish scope

• Published: README.md, .gitattributes, model_hashes.json, fonts/, models/detection/, models/inpainting/, models/ocr/, models/stt/, models/tts/, models/upscaling/, .legacy/models/detection/{magi,magi-v2}/.
• Not published: code/, bin/, .deprecated/, .external_state.json, models/_fixtures/.
• code/ is research and competitor tracking. It stays local and out of the HF model repo.
• .legacy/ holds only curated retired payloads, in the same bucket plus format layout as the main models/.
• models/_fixtures/ is the local place for test fixtures and placeholders. It stays out of HF and currently has no publishable payload.
• HF is an archive: the remote repository keeps payloads that may no longer exist on the development machine. Sync with scripts/maintenance/sync_hf_repo.py --execute --skip-delete unless you deliberately want to prune the remote tree down to the local state.
• model_hashes.json uses schema v2: resources are grouped by bucket path and by asset format (pt, onnx, gguf, native, comfy).