Kiyo

Kiyo is the set of model weights kukkiri runs on: a detector that finds mosaics and a restorer that reconstructs them. The current release is Kiyo v4.

What's in it

Part	Model	Size	What it does
Detector	RF-DETR seg-medium	141 MB, plus a 65 MB fp16 ONNX export	Marks mosaic regions in every frame.
Restorer	BasicVSR++	39 MB	Reconstructs each region from clips of 96 frames at a time.

The files are pickle-free: safetensors for the weights and JSON for the detector's settings. Each one is checked against the SHA-256 listed in kukkiri/weights.py before it is used. See How it works for where each model fits in.

Getting them

The first restore downloads them, and the Windows app does the same on first use. They go to a per-user folder:

Windows: %LOCALAPPDATA%\kukkiri\weights\<version>
Mac and Linux: ~/.cache/kukkiri/weights/<version>

Or to the folder in KUKKIRI_WEIGHTS or --weights. To fetch them ahead of time:

kukkiri download-weights [folder]

On the first run, TensorRT engines are built into the same folder. That takes a few minutes and is cached afterwards.

Versions

Each release is published in its own folder on Hugging Face, and a published folder is never overwritten. A new release downloads into a fresh folder, so an older version stays where it is.

License

The weights are licensed under the same terms as the code: GPL-3.0-only, with an additional permission to combine kukkiri with NVIDIA's CUDA, TensorRT and Video Codec SDK libraries.