Learning Timeline
Key Insights
Resolution Limits
While the model can upscale, native generation quality begins to degrade above 2048x2048 resolution. Attempting 4K generation directly may push VRAM limits and reduce coherence.
Step Count Sweet Spot
Use 8 to 10 steps for the best balance. Lower steps (4-6) result in faster but softer/noisier images. Steps higher than 10 may improve sharpness slightly but can lead to over-sharpening artifacts.
CFG and Negative Prompts
Z-Image Turbo works best with a CFG of 1.0. At this specific value, the model effectively ignores negative prompts. If you need strict text adherence, increase CFG to 1.x or 2.0, but be aware this increases color saturation.
VRAM Requirements
The 11.4 GB model fits within 12 GB VRAM cards. It can run on lower VRAM specs (e.g., 8GB) by utilizing ComfyUI's automatic offloading features.
Step by Step
Installing and Loading the Workflow
- Download the Z-Image JSON workflow file to your local machine.
- Open ComfyUI.
- Click 'Manager' in the main menu.
- Click 'Update ComfyUI' to ensure the latest version is installed.
- Click 'Restart' and then 'Confirm' to apply updates.
- Download the Z-Image Model file (11.4 GB).
- Move the Z-Image Model file into the `ComfyUI/models/diffusion models` directory.
- Download the `AE.safetensors` file.
- Move the `AE.safetensors` file into the `ComfyUI/models/VAE` directory.
- Press 'R' on your keyboard to refresh the ComfyUI model list.
- Drag and drop the downloaded JSON workflow file directly onto the ComfyUI interface to load the node structure.
Configuring and Generating Images
- Locate the resolution node and define width and height (Recommended max: 2048x2048).
- Set the 'Batch Size' to the number of images you wish to generate per run.
- Enter your descriptive text into the prompt text field.
- Locate the Sampler/Generation node settings.
- Set 'Steps' to a value between 8 and 10.
- Set 'CFG' (Guidance Scale) to approximately 1.0.
- Confirm the Sampler is set to 'Euler' (Oiler) with 'DDIM Uniform' (or leave as default 'res multistep').
- Click 'Queue Prompt' (Run) to start generation.
- Wait for the process to complete (approx. 14 seconds on 16GB VRAM).
- Retrieve the final image from the 'Save Image' node preview or the ComfyUI output folder.
