moebius / README.md
jennyxue's picture
Model files
316014b
|
Raw History Blame Contribute Delete
4.17 kB
metadata
license: apache-2.0
base_model:
  - hustvl/Moebius
library_name: mlx
pipeline_tag: image-to-image
tags:
  - mlx
  - moonbox
  - inpainting

Moebius for Moonbox

The files the Moonbox desktop app downloads for these models:

  • Moebius (HUST Vision Lab): Removes objects and repairs damage; everything outside the brushed area stays as it was.

They are the Moebius inpainting UNets (General, Scenes, Faces) and their VAE converted to MLX and Core ML for Apple silicon. Moonbox checks every file against a signed list before using it. You do not need this repository to use Moonbox, and it is not an official release of the original models.

What is here

Path What From
packs/moebius-general/unet.safetensors, packs/moebius-general/temb.npz, packs/moebius-general/manifest.json Moebius 0.22B inpainting UNet, variant 'pretrained' (General) https://huggingface.co/hustvl/Moebius (pretrained/)
packs/moebius-places2/ Moebius fine-tune 'ft_places2' (Scenes) https://huggingface.co/hustvl/Moebius (ft_places2/)
packs/moebius-ffhq/ Moebius fine-tune 'ft_ffhq' (Faces) https://huggingface.co/hustvl/Moebius (ft_ffhq/)
packs/moebius-general/vae.safetensors, packs/moebius-general/coreml/ PixelHacker VAE (frozen SDXL VAE), fp16 + Core ML https://huggingface.co/hustvl/PixelHacker (vae/)
models/ort/ the three UNets and the VAE as ONNX graphs, for the Windows engine https://huggingface.co/hustvl/Moebius, https://huggingface.co/hustvl/PixelHacker
models/hf/BiRefNet/ BiRefNet, used to select a picture's subject (Windows engine) https://huggingface.co/ZhengPeng7/BiRefNet
models/hf/sam2.1-hiera-tiny/ SAM 2.1 Hiera Tiny, used to select objects (Windows engine) https://huggingface.co/facebook/sam2.1-hiera-tiny
models/torch/maskrcnn_resnet50_fpn_v2_coco.safetensors torchvision's Mask R-CNN (ResNet-50 FPN v2, COCO) weights as safetensors (Windows engine) https://github.com/pytorch/vision

Every part was converted, quantized or repackaged from its source unless its row says "as published": treat none of these files as the original release.

Licences

The Scenes and Faces fine-tunes were trained by their authors on Places2 and FFHQ, datasets released for research and non-commercial use; BiRefNet's general weights were trained on DIS5K, also non-commercial. The weights carry the licences below; whether the training data matters for your use is yours to judge.

Part Licence Notice
Moebius 0.22B inpainting UNet, variant 'pretrained' (General) Apache License 2.0 Copyright 2024 Moebius Authors
Moebius fine-tune 'ft_places2' (Scenes) Apache License 2.0 Copyright 2024 Moebius Authors
Moebius fine-tune 'ft_ffhq' (Faces) Apache License 2.0 Copyright 2024 Moebius Authors
PixelHacker VAE (frozen SDXL VAE), fp16 + Core ML MIT License PixelHacker authors (HUST Vision Lab); the SDXL VAE is Copyright (c) Stability AI, MIT
the three UNets and the VAE as ONNX graphs, for the Windows engine Apache License 2.0; MIT License
BiRefNet, used to select a picture's subject (Windows engine) MIT License Copyright (c) 2024 ZhengPeng
SAM 2.1 Hiera Tiny, used to select objects (Windows engine) Apache License 2.0 Copyright (c) Meta Platforms, Inc. and affiliates
torchvision's Mask R-CNN (ResNet-50 FPN v2, COCO) weights as safetensors (Windows engine) BSD 3-Clause License Copyright (c) Soumith Chintala 2016, all rights reserved

The full text of each licence is in LICENSE. The models are the work of their authors, linked above. Moonbox converted and packaged them and claims no rights in them. Use is subject to each licence above.