Learning Timeline
Key Insights
Low VRAM Capability
Using the 'Q3 small GGUF' model (approx. 3.79 GB) allows Z-Image Turbo to run comfortably on GPUs with as little as 4 GB of VRAM.
Matching Quantization Levels
For optimal compatibility and performance, attempt to match the quantization level of the text encoder (e.g., Q4 Medium) with the quantization level of the main GGUF model.
Negative Prompt Usage
Z-Image Turbo generally does not require a negative prompt to function correctly, though the input field remains available in the workflow.
Step by Step
Downloading and Installing Models
- Navigate to the 'Files and versions' tab on the hosted model repository.
- Select a GGUF model file based on available VRAM (e.g., download 'Q3 small GGUF' for 4GB VRAM or 'Q4 medium' for higher quality).
- Move the downloaded GGUF file into the directory: `ComfyUI/models/unet`.
- Navigate to the Quantized Text Encoder repository.
- Download a text encoder that matches the main model's quantization level (e.g., if using Q4 medium model, download the Q4 medium text encoder).
- Move the text encoder file into the directory: `ComfyUI/models/text_encoders`.
- Download the Z-Image GGUF example workflow JSON file to a local drive.
Configuring the Workflow in ComfyUI
- Open the ComfyUI interface in a web browser.
- Drag and drop the downloaded Z-Image GGUF workflow JSON file directly onto the canvas.
- Check for nodes highlighted in red. If present, click 'Manager' in the menu.
- Click 'Install Missing Custom Nodes' (or browse Custom Nodes for GGUF support) and install the required extensions.
- Click 'Restart' to reload ComfyUI with the new nodes active.
- Locate the GGUF U-Net loader node.
- Click the dropdown menu and select the GGUF model file placed in the `unet` folder earlier.
- Locate the CLIP/Text Encoder loader node.
- Click the dropdown menu and select the quantized text encoder file placed in the `text_encoders` folder.
- Input the desired image description into the 'Positive Prompt' text box.
- Set the 'Width' and 'Height' values in the empty latent image node (e.g., 1024 for both).
- Click 'Queue Prompt' to start the generation.
