vanch007/MOSS-Image-SFT-mlx-bf16

MOSS-Image-SFT (BF16) for Apple Silicon (MLX / mflux)

Native Apple Silicon implementation of MOSS-Image-SFT ported to Apple MLX via mflux.

• Precision: BF16 (Full Precision Baseline)
• Model Size: Transformer 9.52 GB
• Architecture: 42-layer MOSS-Image DiT Transformer + Qwen 3.5 Text Encoder + FLUX.2 Small VAE
• Native Apple Silicon Support: Metal Performance Shaders (MPS) & Unified Memory

3-Precision Comprehensive Benchmark (Tested on Apple M3 Max)

Precision	Transformer VRAM	Compression	1024² Step Time	Fidelity	Engineering Recommendation
BF16	9.52 GB	Baseline (100%)	~13.9 s / step	⭐⭐⭐⭐⭐ 100%	Master production / Highest detail
8-bit	5.03 GB	-47.2%	~17.1 s / step	⭐⭐⭐⭐⭐ 99.8%	👑 Recommended for Daily Production
4-bit	2.66 GB	-72.1%	~20.9 s / step	⭐⭐⭐⭐ 92.0%	16 GB unified memory devices

Precision Comparison Cards across 4 Scenarios

Scenario 1: Realistic Fluffy Cat with Golden Hour Backlighting (1024x1024)

Scenario 2: Oriental Classical Portrait & Skin Texture (1024x1024)

Scenario 3: Chinese Calligraphy & Live Streaming UI Layout (944x1680)

Scenario 4: Macro Precision Watchmaking Workbench Still Life (1024x1024)

Usage with mflux

Installation

git clone -b feature/moss-image-sft https://github.com/vanch007/mflux.git
cd mflux
pip install -e .

CLI Generation

# Generate with BF16 model
mflux-generate-moss-image --model vanch007/MOSS-Image-SFT-mlx-bf16 \
    --prompt "A majestic fluffy cat sitting on a wooden bench, golden hour lighting, 8k photo" \
    --steps 12 --guidance 3.0 --width 1024 --height 1024 --seed 42 --output output.png

Python API

from mflux.models.moss_image.variants.txt2img.moss_image import MOSSImage

model = MOSSImage(model_path="vanch007/MOSS-Image-SFT-mlx-bf16")
image = model.generate_image(
    prompt="A majestic fluffy cat sitting on a wooden bench, golden hour lighting, 8k photo",
    width=1024,
    height=1024,
    num_inference_steps=12,
    guidance_scale=3.0,
    seed=42
)
image.save("moss_cat.png")

Acknowledgements

• OpenMOSS Team for the original MOSS-Image model.
• mflux-community for the Apple Silicon MLX framework.