---
license: apache-2.0
library_name: mlx
pipeline_tag: image-to-image
base_model: hustvl/Moebius
tags:
  - mlx
  - image-inpainting
  - inpainting
  - diffusion
  - moebius
  - ft_ffhq
  - mxfp4
---

# Moebius ft_ffhq MLX mxfp4

This folder contains a converted MLX version of the **ft_ffhq** Moebius checkpoint in **mxfp4** form. `ft_ffhq` means the Moebius checkpoint fine-tuned on FFHQ, for face and portrait inpainting.

Original upstream model: [hustvl/Moebius](https://huggingface.co/hustvl/Moebius)  
Original source repository: [hustvl/Moebius](https://github.com/hustvl/Moebius)

Moebius checkpoint fine-tuned on FFHQ. This is another face and portrait-oriented fine-tune.

## Identity

| Field | Value |
|---|---|
| Variant name | `ft_ffhq-mxfp4` |
| Checkpoint meaning | Fine-tuned on FFHQ for face and portrait inpainting. |
| Original Moebius checkpoint family | `ft_ffhq` |
| Original checkpoint type | FFHQ fine-tune |
| Source PyTorch checkpoint | `Moebius-Models/ft_ffhq/diffusion_pytorch_model.bin` |
| MLX precision / quantization label | `mxfp4` |
| Image size | 512 x 512 |
| Latent size | 64 x 64 |
| Latent channels | 4 |
| Mask channels | 1 |
| Conditioning IDs | 20 |
| VAE scaling factor | 0.13025 |
| Noise offset | 0.0357 |

## Quantization

MLX MXFP4-style 4-bit export. The manifest selects `unet_quantized.safetensors`; VAE encoder and decoder remain regular f16 safetensors. Runtime quantization metadata uses `mode: mxfp4` with group size 32.

- UNet precision mode: `mxfp4`.
- MLX grouped quantization is used for supported linear layers; grouped quantized weights are loaded through the Moebius-MLX manifest/runtime.
- Quantization config: 4 bits, group size 32, `mxfp4` mode.
- The VAE encoder and decoder remain regular f16 safetensors for this variant.

## Manifest-selected deployment files

These are the files selected by `manifest.json` when the Moebius-MLX runtime loads this variant.

| Component | File | Size |
|---|---|---:|
| UNet | `unet_quantized.safetensors` | 421.75 MB |
| VAE encoder | `vae_encoder.safetensors` | 68.34 MB |
| VAE decoder | `vae_decoder.safetensors` | 99.00 MB |

## Files in this folder

- `unet.safetensors`
- `unet_quantized.safetensors` (selected by manifest)
- `vae_decoder.safetensors` (selected by manifest)
- `vae_encoder.safetensors` (selected by manifest)
- `manifest.json` (runtime metadata and file selection)

A minimal runtime package needs `manifest.json` and the manifest-selected files above. Extra source or fallback files are optional and are not required for inference.

## Runtime expectations

This is not a Transformers or Diffusers-native checkpoint. It is intended for the Swift/MLX runtime in [Moebius-MLX](https://github.com/kylehowells/Moebius-MLX). The runtime reads `manifest.json`, loads the selected safetensors files, builds the Moebius UNet and VAE modules, and runs the DDIM inpainting pipeline.

Pipeline constants must match the manifest:

- DDIM scheduler: `scaled_linear`, beta start 0.00085, beta end 0.012, 1000 train timesteps, clip sample false
- 512 x 512 image resolution and 64 x 64 latent resolution
- 9-channel UNet input: noisy latent, mask, and masked-image latent
- VAE scaling factor 0.13025


## Attribution

Moebius was released by the original authors as [hustvl/Moebius](https://huggingface.co/hustvl/Moebius). This folder is a format conversion and/or quantized MLX packaging of the original PyTorch weights, not a newly trained model.