vanch007/Qwen-Image-2.1-Heretic-mlx-8bit

Qwen-Image-2.1-Heretic (8-bit) — MLX Native

Native Apple Silicon (MLX) port of Qwen-Image-2.1 with the Heretic text encoder (pottokao/Qwen-Image-2.1-Text-Encoder-Heretic).

• Precision: 8-bit
• Size on disk: ~16.6 GB
• Features: Native Apple Silicon Metal acceleration, Text-to-Image, Multi-reference Image Editing, RGBA transparent output, optional Viggle Turbo LoRA acceleration.
• Description: Balanced 8-bit quantized MLX port preserving 95%+ of original detail at roughly half the disk and memory footprint.

Requirements

Install mflux:

pip install mflux

Usage

CLI

mflux-generate-qwen-2.1-edit --model vanch007/Qwen-Image-2.1-Heretic-mlx-8bit --prompt "A serene mountain lake at golden hour" --steps 40 --seed 42 --output lake.png

Turbo Acceleration (6 steps)

mflux-generate-qwen-2.1-edit --model vanch007/Qwen-Image-2.1-Heretic-mlx-8bit --lora-preset viggle-turbo --prompt "A red ceramic teapot on a wooden table" --steps 6 --guidance 1.0 --output turbo.png