ComfyUI Z Image Turbo: generation and upscaling tutorial

This tutorial walks through a ComfyUI workflow for Z Image Turbo, covering image generation, ControlNet guidance and image-to-image upscaling. The presenter describes the 6B model as suitable for lower-end GPUs and reports an 11-second sampling run in the example. The transcript does not specify the hardware, so that timing is not a general performance benchmark.

The core workflow loads the model, text encoder and VAE, then applies optional LoRAs and a positive prompt before sampling and decoding. The negative prompt is unused in this setup. With denoise set to 1.0, the presenter generates a fresh image rather than preserving the loaded input.

ControlNet adds reference guidance through Canny edges, pose and depth, each with an adjustable strength. The first demonstration uses pose guidance while setting edge and depth strengths to zero.

For upscaling, the presenter reloads a generated image as both input and control reference, selects a scale factor of 2 and lowers denoise to 0.4. LoRAs are disabled to avoid imposing their style, while edge and depth guidance help retain the original structure. The examples also show the limits: sampling can introduce new details, and dark images may yield few detected edges unless the threshold changes.