Z-Image LoRA Trainer

Train custom LoRA models for Z-Image Turbo. Upload your training images and configure the settings below.

Important: This Space requires an L4 GPU ($0.80/hr) for training. Upgrade in Settings before starting.

GPU Status

Upload 6-20 training images

Tips for best results:

Trigger Word

Use a unique word that doesn't exist in normal vocabulary

Output Name

Name for your LoRA file

Training Steps

More steps = better quality but longer training

500 5000

Learning Rate

0.00001 0.0005

LoRA Rank

Higher = more capacity but more VRAM

4 32

Training Resolution

Lower = faster and less VRAM

256 1024

Batch Size

Keep at 1 for L4 GPU

1 4

Download LoRA

Training Log

Upload Images: Drag and drop 6-20 training images
Set Trigger Word: Choose a unique trigger word (e.g., <mychar>)
Configure Settings: Adjust training parameters (defaults work well for most cases)
Start Training: Click the button and wait (1-2 hours for 2000 steps)
Download: Get your LoRA file when training completes

Use your LoRA with ComfyUI or diffusers:

pipe.load_lora_weights("your_lora.safetensors")
image = pipe("<trigger_word>, your prompt here").images[0]