ComfyUI image-to-image setup with Z-Image Turbo

This tutorial adapts an existing ComfyUI text-to-image workflow to accept a reference image. It assumes the starting workflow already works; it does not cover installation. The example uses Z-Image Turbo, and the speaker notes that other diffusion models may need a slightly different setup.

The main change replaces Empty Latent Image with Load Image and VAE Encode. An image output cannot connect directly to the KSampler's latent input because they carry different data types. VAE Encode converts the input into latent space, while VAE Decode converts the generated result back into a visible image. Both nodes use the same loaded VAE model.

The speaker adds a resize node before encoding and recommends preserving the input's aspect ratio. He describes roughly 1024 by 1024 pixels as Z-Image Turbo's preferred size and warns that a much larger input, such as a 3840 by 2160 photo, may exceed the machine's resources or produce distorted results.

Denoise controls how far the output departs from the reference. The speaker suggests 0.2 to 0.4 for subtle edits and 0.5 to 0.7 for more noticeable changes. At higher values, the original becomes a looser guide; he describes 1.0 as effectively discarding it. These are practical guidelines rather than guarantees. The closing explanation stresses the prompt's role and points to ControlNet or similar additions for retaining poses or style.