Update README.md

README.md

@@ -9,6 +9,7 @@ This model is a LoRA used to fix the acceleration capability of Z-Image Turbo Lo
 
 LoRAs trained directly based on Z-Image Turbo will lose their acceleration capability. Images generated under acceleration configuration (steps=8, cfg=1) become blurry, while images generated under non-acceleration configuration (steps=30, cfg=2) remain normal.
 
+
 ## Results
 
 Training Data:
@@ -21,6 +22,20 @@ Generation Results:
 |-|-|-|
 |![](assets/image_base_acc.jpg)|![](assets/image_base_nonacc.jpg)|![](assets/image_with_our_lora.jpg)|
 
+## Training with Z-Image Turbo
+
+If you want to train LoRAs based on Z-Image Turbo while maintaining its acceleration capability, please refer to our detailed training strategies guide:
+
+📖 [**Training Strategies of Z-Image Turbo**](https://huggingface.co/blog/kelseye/training-strategies-of-z-image-turbo)
+
+This guide covers four different training approaches:
+- **Scheme 1**: Standard SFT Training + No Acceleration Configuration
+- **Scheme 2**: Differential LoRA Training + Acceleration Configuration
+- **Scheme 3**: Standard SFT + Trajectory Imitation Distillation
+- **Scheme 4**: Standard SFT + Loading DistillPatch LoRA (Recommended)
+
+We recommend **Scheme 4** as it offers the best trade-off between training simplicity and inference speed.
+
 ## Inference Code
 
 ```python