zimengxiong/Qwen-Image-Layered-6bit

Qwen-Image-Layered (6-bit Quantized for MFLUX)

6-bit quantized version of Qwen/Qwen-Image-Layered optimized for Apple Silicon via MFLUX.

Version	Size	VRAM
BF16 (original)	~55 GB	~55 GB
6-bit (this)	~31 GB	~29 GB

What is Qwen-Image-Layered?

Qwen-Image-Layered decomposes any image into multiple RGBA layers with transparency. Each layer can be independently edited, moved, resized, or recolored—enabling high-fidelity, consistent image editing.

Quick Start

Installation

Requires MFLUX with Qwen-Image-Layered support (PR #302):

# Install from the PR
git clone https://github.com/zimengxiong/mflux.git
cd mflux
uv sync

Usage

uv run flux-generate-qwen-layered \
  --image input.png \
  --layers 4 \
  --steps 50 \
  -q 6 \
  --output-dir ./layers

Output: 4 RGBA PNG files (layer_0.png, layer_1.png, etc.) with transparency.

Parameters

Parameter	Description	Default
--image	Input image path	Required
--layers	Number of layers to decompose	4
--steps	Inference steps	50
-q	Quantization (4, 6, or 8-bit)	None (BF16)
--resolution	Resolution bucket (640 or 1024)	640
--output-dir	Output directory for layers	./

Example

Input image:

Model Architecture

• Transformer: QwenImageTransformer2DModel with Layer3D RoPE (3D positional encoding for [layer, height, width])
• VAE: AutoencoderKLQwenImage (RGBA 4-channel with 3D temporal convolutions)
• Text Encoder: Qwen2.5-VL for conditioning

Hardware Requirements

• Apple Silicon Mac (M1/M2/M3/M4 series)
• Minimum 24GB unified memory (32GB+ recommended)
• macOS 13.0+

Links

• 📦 MFLUX - Fast MLX inference for Apple Silicon
• 🤗 Original Model - Full BF16 weights
• 📑 Paper - Research paper
• 📑 Blog - Official blog post

License

Apache 2.0 (same as original model)

Citation

@misc{yin2025qwenimagelayered,
      title={Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition}, 
      author={Shengming Yin, Zekai Zhang, Zecheng Tang, Kaiyuan Gao, Xiao Xu, Kun Yan, Jiahao Li, Yilei Chen, Yuxiang Chen, Heung-Yeung Shum, Lionel M. Ni, Jingren Zhou, Junyang Lin, Chenfei Wu},
      year={2025},
      eprint={2512.15603},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.15603}, 
}