botservice-dev/parallax-models

Parallax Wallpaper Studio models

ONNX exports of three open models, made to run on the phone's CPU (ONNX Runtime) inside the Android app Parallax Wallpaper Studio. The app downloads them on demand: the depth model for depth-based parallax, and all three for "Split into layers", which turns one photo into a layered wallpaper.

File	Model	Used for	Licence
depth-anything-v2-small-int8.onnx	Depth Anything V2 Small, int8 dynamic quantisation	Depth map	Apache-2.0
vitmatte-small-512-fp16w.onnx	ViTMatte Small (Composition-1k), fixed 512×512 input, fp16 weights	Refining layer edges	Apache-2.0
big-lama-512-fp16w.onnx	LaMa (big-lama), fixed 512×512 input, fp16 weights	Filling the holes behind cut-out layers	Apache-2.0

Notes on the exports:

• The depth model matches the full-precision model to 0.9998 correlation.
• LaMa's FFTs have no ONNX export, so they are rewritten as constant DFT matrix multiplications. That fixes the input size at 512×512.
• "fp16 weights" means the weights are stored as float16 and the maths runs in float32.

Credits

• Depth Anything V2: Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, Hengshuang Zhao. Depth Anything V2, 2024. Only the Small model is Apache-2.0; Base and Large are not, and are not used here.
• ViTMatte: Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang. ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers, 2023 (hustvl).
• LaMa: Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, Victor Lempitsky. Resolution-robust Large Mask Inpainting with Fourier Convolutions, WACV 2022 (Samsung AI Center).

These are format conversions of the original weights. The licences and copyright of the original works apply.