ComfyUI and Z-Image Turbo: text-to-image workflow tutorial

This tutorial builds a ComfyUI text-to-image workflow from an empty canvas. The presenter explains each connection around KSampler: the diffusion model, positive and negative prompt conditioning, and an empty latent image that sets the output dimensions. The example prompt asks for a sunflower on a meadow.

For Z-Image Turbo, the presenter selects Qwen-3-4B in Load CLIP and sets its type to Lumina-2. The walkthrough stresses that the text encoder and VAE must match the diffusion model; incompatible choices can cause errors or poor results. It also notes that some workflows ignore negative prompts when CFG is set to one. Image dimensions should suit the model, and the speaker warns that attempting 5000-pixel output may stall the machine.

VAE Decode converts KSampler's latent output into an RGB image, using AE.safetensors in this example. Save Image writes the result to the output folder. The presenter explains that the first run takes longer because ComfyUI loads the models into memory, and leaves KSampler settings unchanged to focus on the pipeline rather than output quality.

The final section covers saving the workflow in the interface and exporting a JSON file that another ComfyUI user can drag onto their canvas. The description places the demonstration in Floyo's browser environment, so this is a workflow tutorial rather than a guide to installing local AI software.