ai-toolkit-trainer

Use ostris AI-Toolkit: prepare a dataset of images or short clips with matching .txt captions, create a job in the web UI, and enable Multi-stage training for both high-noise and low-noise experts. Start with rank 16, learning rate 5e-5, and 1500-2500 steps.

Z-Image LoRA training runs on modest consumer GPUs using AdamW8bit optimizer with float8 quantization, batch size 1, and 512-768 resolution. Keep Multi-stage off since Z-Image is a single-stream model, and start with rank 16-32 at learning rate 1e-4.

Yes, the RunPod installer script sets up AI-Toolkit in /workspace/ai-toolkit on the persistent volume and is idempotent. Use the PyTorch 2.8.0 template with a 100GB disk, set AI_TOOLKIT_AUTH, and access the UI on port 8675 via the RunPod proxy URL.

WAN image LoRAs need 24GB+ VRAM locally with quantization, while WAN video LoRAs, high resolution, or high rank jobs are best on cloud GPUs like RTX 5090 or RTX 6000 Pro. Z-Image image LoRAs are the lightest and train on smaller cards with float8.

This error means the venv's Torch stack is mismatched. Activate the AI-Toolkit venv, uninstall torch, torchaudio, and torchvision, then reinstall them together from the correct PyTorch CUDA index URL for your GPU.

Copy the .safetensors file into ComfyUI's models/loras folder and load it with LoraLoaderModelOnly. WAN 2.2 requires applying the LoRA to both HighNoise and LowNoise branches, while Z-Image uses a single loader at strength 0.7-1.0.