Z-Image-Turbo Training Adapter

Introduction

This is the Z-Image-Turbo Training Adapter, a special LoRA designed to help you fine-tune the Tongyi-MAI/Z-Image-Turbo model. It's built to make your training runs smoother by preserving the speed benefits of the "Turbo" model. You can also use it on its own for inference to dial back some of that "Turbo" effect.

How to Use

• Trigger word: No trigger word is needed.
• Purpose: This adapter is primarily for use during training when you're fine-tuning Z-Image-Turbo. It helps prevent the model's distillation (its speed) from breaking down too quickly.
• Inference: For inference, you'd typically remove this adapter after your fine-tuning is complete. This lets your newly trained information run at the original distilled speeds. However, you can also use it as a general de-distillation LoRA if you want to remove some of the "Turbo" characteristics during inference.
• Training with it: When fine-tuning Z-Image-Turbo, train your new LoRA on top of this adapter. This way, your new LoRA only learns your subject, and the distillation stays intact.
• Limitations: It's best for shorter training runs like styles, concepts, or characters. Long training runs might still cause the distillation to break down, potentially leading to artifacts if you remove the adapter later.

FAQ

Q: What's the trigger word?

A: No trigger word is required for the Z-Image-Turbo Training Adapter.

Q: Why do I need this adapter for training?

A: When you train directly on a "Turbo" (step-distilled) model, its speed benefits can quickly break down. This adapter slows that process down significantly, letting you do shorter training runs while keeping the model fast.

Q: Can I use this for long training runs?

A: It's best for shorter runs (styles, concepts, characters). For very long training sessions, the distillation might still break down over time, which could lead to artifacts if you remove the adapter for inference.

Q: How does it help my new LoRA?

A: By training your new LoRA on top of this adapter, you're essentially de-distilling the base model first. This means your new LoRA only learns your specific subject, and when you remove this adapter for inference, your new information runs at the original distilled speeds.

Q: What LoRA strength should I use?

A: The documentation doesn't specify a recommended LoRA strength for using this adapter. It's designed to be present during training and then removed for inference. If using it as a de-distillation LoRA for inference, experiment to find what works best for your desired effect.

Technical Details

• Base model: Tongyi-MAI/Z-Image-Turbo
• Training info: The adapter was created by generating thousands of images at various sizes and aspect ratios using Tongyi-MAI/Z-Image-Turbo. A LoRA was then trained on these images at a low learning rate (1e-5) to break down the distillation while preserving the model's existing knowledge.
• Source: Hugging Face