transparent-models

Model checkpoints for transparent, a local, GPU-accelerated background-removal and image-editing toolkit for Linux. The app downloads these on demand and verifies each file against the SHA-256 recorded in its model catalog.

All checkpoints are F16 GGUF conversions for inference with vision.cpp (ggml, Vulkan/CPU). Stored via Git LFS.

Denoising (SCUNet)

Swin-Conv-UNet blind real-world color denoising (cszn/SCUNet, Zhang et al. 2022, Apache-2.0). Converted from the upstream .pth checkpoints by the vision.cpp fork's scripts/convert.py scunet. 1:1 resolution (no upscaling). Large images are tiled internally.

File	Variant	Size	Notes
scunet-color-real-gan-F16.gguf	real GAN	~36 MB	Sharper results, occasional artifacts. App default.
scunet-color-real-psnr-F16.gguf	real PSNR	~36 MB	Smoother, cleaner metrics.

Integrity (SHA-256, also recorded in the app's ModelCatalog):

692e7b7d24979a1fc2ca15f2984ded55ff7b6f7bf24882a7c9aa81a03a0c9a7e  scunet-color-real-gan-F16.gguf
e839ee6839f265b254d498510bdb3c2e62700b8b34cfe8679408511b5c4c5254  scunet-color-real-psnr-F16.gguf

Segmentation (FeyNoBg)

BiRefNet with a deepened Swin-L stage 3 (feyninc/FeyNobg, Apache-2.0). Converted from the upstream .safetensors by the vision.cpp fork's scripts/convert.py birefnet, which maps the timm-style backbone keys and records the non-default stage depths in the GGUF.

File	Size	Notes
FeyNoBg-F16.gguf	~526 MB	Fixed 1024x1024 inference.

Integrity (SHA-256, also recorded in the app's ModelCatalog):

110854aec5f093a630b31722b5d633d132bc1f4233a1976d80c304df9a566f89  FeyNoBg-F16.gguf

Adding models

Convert (or obtain) the GGUF, place it in this directory, and add an LFS rule for its extension if not already tracked:

git lfs track "*.gguf"
git add file.gguf .gitattributes

Then register the model in the app's ModelCatalog (src/core/ModelCatalog.cpp) with its filename, license, size, download URL and SHA-256. Models are fetched from this repo's raw LFS URLs: .../media/branch/main/<file>.