Model card and the original license

README.md

@@ -0,0 +1,80 @@
+---
+base_model: hustvl/Moebius
+base_model_relation: quantized
+license: mit
+pipeline_tag: image-to-image
+library_name: mlx
+tags:
+- mlx
+- moonbox
+- moebius
+- inpainting
+- object-removal
+---
+
+# Moebius MLX
+
+MLX builds of [Moebius](https://huggingface.co/hustvl/Moebius), made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's
+files together. Only the parts listed below are converted: the rest of the pipeline comes from the original
+repository unless the section below says otherwise.
+
+## Files
+
+The last column is how far each build is from the original PyTorch model on the same inputs (see the section below
+for what was measured).
+
+| Precision | File | Size | sha256 | Checked against the original |
+|---|---|---|---|---|
+| fp16 | [`fp16/temb.npz`](https://huggingface.co/moonbox-ai/Moebius-MLX/blob/main/fp16/temb.npz) | 5.1 MB | `b5b4a28cb0ba` | PSNR 41.0 dB mean / 33.9 worst, LPIPS <= 0.008 |
+| fp16 | [`fp16/unet.safetensors`](https://huggingface.co/moonbox-ai/Moebius-MLX/blob/main/fp16/unet.safetensors) | 476.8 MB | `e75b9e1549e4` | PSNR 41.0 dB mean / 33.9 worst, LPIPS <= 0.008 |
+| fp16 | [`fp16/vae.safetensors`](https://huggingface.co/moonbox-ai/Moebius-MLX/blob/main/fp16/vae.safetensors) | 167.3 MB | `81593a7667f4` | PSNR 41.0 dB mean / 33.9 worst, LPIPS <= 0.008 |
+
+## How to run
+
+Moebius (the `pretrained` general model) with the PixelHacker VAE, converted for the MLX port in Moonbox's Moebius
+engine. Each folder holds:
+
+| File | What |
+|---|---|
+| `unet.safetensors` | the 0.22 B UNet: BatchNorms, norm affines and the constant conditioning folded into the weights; 1x1 convs and attention projections stored as `[in, out]` matrices (MLX affine-quantized, group 64, in the `8bit` / `6bit` / `4bit` folders: `<name>.q`, `.scales`, `.biases`) |
+| `vae.safetensors` | PixelHacker's VAE (fp16 in the quantized folders) |
+| `temb.npz` | `silu(time_embedding(t))` for t = 0..999 |
+
+The layout is the one `moebius_od/mlx_unet.py` and `mlx_vae.py` in Moonbox read (NHWC, 512x512, 20 DDIM steps, CFG 2,
+as upstream's README). That runtime isn't part of this repository.
+
+**How they were checked.** Through the same MLX engine against the float32 weights, removing the main subject of 8
+photos (seed 0): PSNR and LPIPS of the 512 px output. Pass: LPIPS <= 0.15 on every image. The MLX port itself was
+checked against the official PyTorch pipeline when the engine was written.
+
+| Precision | Size | PSNR vs float32 (mean / worst) | LPIPS (worst) | Status |
+|---|---|---|---|---|
+| `fp16/` | 649 MB | 41.0 / 33.9 dB | 0.008 | passes |
+| `bf16/` | 649 MB | 32.2 / 27.2 dB | 0.050 | passes |
+| `8bit/` | 495 MB | 37.4 / 33.0 dB | 0.022 | passes |
+| `6bit/` | 453 MB | 32.9 / 26.4 dB | 0.066 | passes |
+| `4bit/` | 412 MB | 22.4 / 16.7 dB | 0.380 | **below threshold, experimental** |
+
+About 4-5 s per inpainting (20 steps, 512 px) on an Apple M5.
+
+## How these were made
+
+From [Moebius](https://huggingface.co/hustvl/Moebius) at revision `cd01f47fb6`: the original weights loaded in MLX and stored as bf16 or fp16, or
+quantized with MLX's affine quantization (`8bit`, `6bit`, `4bit`, group size 64; layers MLX can't quantize, such as
+convolutions and norms, stay bf16). Every build was checked against the original before upload; the `sha256` column
+above is what the files must hash to.
+
+## License
+
+Apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this
+repository, and it governs your use of these files. The original model card follows unchanged.
+
+---
+
+# Original model card
+
+《Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance》  
+Project Page: https://hustvl.github.io/Moebius  
+Paper: https://arxiv.org/abs/2606.19195  
+Code: https://github.com/hustvl/Moebius  
+weight: https://huggingface.co/hustvl/Moebius/tree/main