ComfyUI: set up Z-Image Turbo models and loaders

This tutorial explains ComfyUI model setup using Z-Image Turbo. It compares an all-in-one checkpoint with separate diffusion model, text encoder and VAE files, then connects each loader to an existing image-generation workflow.

The download example uses an INT8 diffusion model of about 6 GB, a Qwen 3 4B FP8 mixed text encoder and ae.safetensors. The files go into models/diffusion_models, models/text_encoders and models/vae. The optional checkpoint, about 10 GB, belongs in models/checkpoints. The presenter explains filename labels such as BF16, FP8 and INT8 as choices that affect file size and memory use. GGUF variants need an extra custom node, which this lesson does not cover.

For the separate files, the walkthrough uses Load Diffusion Model, Load CLIP and Load VAE. It selects Lumina 2 as the text encoder type for Z-Image and leaves device settings at default. The presenter notes that putting the encoder on CPU saves video memory but slows prompt processing.

The comparison uses the same prompt and seed. The presenter attributes small image differences to the checkpoint's FP8 diffusion model versus the separate INT8 model. Troubleshooting covers refreshing lists with R, checking folders, reconnecting a missing VAE wire and correcting an incompatible Flux 2 VAE. Practice and completed workflows accompany the lesson.