ERNIE-Image-Turbo-MLX-4bit

This repository contains the 4-bit quantized version of Baidu's ERNIE-Image-Turbo text-to-image model, optimized for Apple Silicon using MLX.

ERNIE-Image-Turbo is a fast, high-quality 12B parameter text-to-image transformer. Quantized to 4-bit, the model size is reduced to approximately 6.5 GB, allowing it to run smoothly on Macs with 8GB or 16GB of unified memory without triggering Metal GPU watchdog timeouts.

Model Details

• Original Model: baidu/ERNIE-Image-Turbo
• MindCraft Studio: MindCraft Studio Homepage
• Quantization: 4-bit weight-only quantization (quantize=4) via MLX.
• Inference Steps: Recommended 8 steps.
• Guidance Scale: 1.0 (CFG is disabled/forced to 1.0 for the Turbo architecture).

How to Use in MindCraft Studio

1. Open MindCraft Studio.
2. Go to Settings -> Model Library -> Import Local Model.
3. Point to this directory.
4. The model will show up as ERNIE-IMAGE (Turbo) and is ready for high-speed generation.

Credits & References

• Original Model: baidu/ERNIE-Image-Turbo
• MindCraft Studio: themindstudio.cc/mindcraft
• MLX Framework: Apple Machine Learning Research