Z-image Base: simple gen & inpainting workflows for high quality realism + tips & info — zimage_base_simple

Use the simple Z-image Base workflow to turn text prompts into high-resolution images; a separate inpainting workflow uses an existing image and mask for image-to-image generation, localized edits, and inpainting, producing edited images.

The described image-generation flow uses two sampler passes: the first does most of the work, and a low-noise second pass adds fine detail.

The updated inpainting workflow removes its second KSampler and blends the original image with the edited area through a semi-transparent mask to help reduce visible seams.

For a starting prompt, describe the subject and style first, then add relevant scene, background, composition, and lighting details.

Suggestion · non vérifié

The author describes the generated images as sharp and suitable for direct upscaling.

For the separate inpainting workflow, the setup notes call for ComfyUI_essentials and ComfyUI LayerStyle custom nodes.

The simple workflow names these model files: ae.safetensors (VAE), qwen_3_4b.safetensors (text encoder), and z-image-BF16.gguf (UNet).

The simple workflow lists RES4LYF, rgthree-comfy, and ComfyUI-GGUF; the notes describe RGTHREE and ComfyUI GGUF as optional.

The suggested two-stage settings are res_2s with beta, 22 steps, and 1.00 denoise, followed by res_2s with normal, 3 steps, and 0.15 denoise.

Try CFG 4 as a starting point; the notes also give values up to CFG 7 for different results.

Suggestion · non vérifié

Preset sizes are grouped by aspect ratio and divisible by 16; changing size or aspect ratio can affect composition, detail, and lighting.

The page details list no model-use license; check the applicable terms before using the model.

Suggestion · non vérifié

The notes warn that high-resolution generation can take a long time and sizes above 1920 pixels per side may look odd.