When training LoRAs with AI-Toolkit, you usually need the original base model files stored locally. If you don’t store them locally, ai-toolkit might download the same files multiple times onto temp directory. Also, downloading models via the browser is slow and unreliable for large repositories, so the recommended approach is using the hf command-line tool from the huggingface_hub Python module.
If you’re thinking about purchasing a new GPU, we’d greatly appreciate it if you used our Amazon Associate links. The price you pay will be exactly the same, but Amazon provides us with a small commission for each purchase. It’s a simple way to support our site and helps us keep creating useful content for you. Recommended GPUs: RTX 5090, RTX 5080, and RTX 5070. #ad
This guide shows how to download popular models like Z-Image-Turbo and Qwen-Image-2512 specifically for LoRA training.
1. Install the Hugging Face CLI
Install or upgrade the Hugging Face Hub client:
Verify the CLI is available:
2. Authenticate (Recommended)
Some models are gated, and authentication also improves download reliability. You will need a read-only token. To get a read-only Hugging Face token, go to your Hugging Face account settings and open the Access Tokens page. Create a new token, set the token type to Read, generate it, and copy the token somewhere safe. Note that you can only view the token once. This token can then be used when logging in with hf auth login to download models.
To confirm login:
3. Download the Base Model Repository
AI-Toolkit expects local model files. Always download into a dedicated folder.
Example: Z-Image-Turbo (for LoRA training)
This ensures:
-
All weights are downloaded locally
-
Files are safe to move or back up
-
AI-Toolkit can read them directly
Example: Qwen-Image-2512
4. Verify Files Before Training
After downloading, your model folder should contain:
-
.safetensorsor weight files -
configuration files (
config.json, etc.) -
tokenizer / text encoder files (if required)
Point AI-Toolkit’s base_model_path to this folder when starting LoRA training.
Conclusion
For LoRA training with AI-Toolkit, the hf CLI is the safest and fastest way to download Hugging Face models locally. Whether you’re training on Z-Image-Turbo, Qwen-Image-2512, or other modern image models, this approach ensures your base model is clean, portable, and ready for training.
Reference
Further Reading
Generate Realistic Images with Z-Image-Turbo GGUF in ComfyUI
Leave a Reply