Z Image Inpaint II

2026年3月21日 20:14

【なでしこジャパンが2大会ぶり優勝｜日本×オーストラリア｜決勝ハイライト】AFC女子アジアカップオーストラリア2026

Brave Matildas fall agonisingly short | Japan win 2026 Women's Asian Cup Final | Highlights

【チェッカーズ Best Performance on NHK】レッツゴーヤング　冬号 Vol.1Z Image Inpaint II Workflow Guide

【チェッカーズ Best Performance on NHK】レッツゴーヤング　冬号 Vol.2

A streamlined Z Image inpainting workflow using ZImageFunControlnet for ControlNet-based inpainting. Compared to the full version, this is a simplified pipeline without LoRA stacking, memory management nodes, translation, or SeedVR2 upscaling — ideal for sharing and distribution.

Overall Flow Overview

[HSWQZImage FP8 Loader] → [SageAttention] → [TorchSettings] → [EasyCache] → [ZImageFunControlnet] → [DiffDiffusionAdv] → [AuraFlow Sampling] → [KSampler]
                                                                      ↑               ↑          ↑                                                      ↓
                                                    [ModelPatchLoader] ┘               |          |                                              [VAEDecodeTiled]
                                                                          [VAELoader] ─┤          |                                                      ↓
                                                                                       |          |                                               [SaveImage]
                                                            [LoadImage] → [InpaintPreprocessor] → ┤
                                                                     └──→ [VAEEncode] ────────────┘
                                                                     └──→ mask ───────────────────┘

[CLIPLoader] → [CLIPTextEncode] → positive → [KSampler]
                       └──→ [ConditioningZeroOut] → negative ↗

Detailed Stage Breakdown

1. Model Loading Pipeline

2. ControlNet Inpainting (ZImageFunControlnet)

This is the core difference from the first workflow. Instead of InpaintModelConditioning, this workflow uses a ControlNet-based approach.

Connection details:

model ← EasyCache output (base Z Image model)
model_patch ← ModelPatchLoaderCustom (ControlNet weights)
vae ← VAELoader
inpaint_image ← InpaintPreprocessor output
mask ← LoadImage mask output
image input is not connected (unused for inpainting mode)

3. CLIP & Text Processing

NOTE: Unlike the full version, there is no translation switch or text preview node. Enter prompts directly in English.

4. Differential Diffusion & Sampling

5. Sampling Settings

IMPORTANT: denoise=1 here because the ControlNet and DifferentialDiffusion handle the blending — the mask determines what changes, not the denoise value.

6. Decoding & Output

No memory management or upscale nodes in this version. Single output image.

Key Differences from Z Image Inpaint I

Usage Instructions

Basic Usage

Prepare Image & Mask
- Load image into the LoadImage node.
- Use Open Mask Editor to paint the area to inpaint.
Input Prompt
- Enter prompt directly in CLIPTextEncode in English.
Adjust Parameters
- ZImageFunControlnet strength: Adjust ControlNet influence (default 1).
- DifferentialDiffusionAdvanced threshold: Controls mask expansion (default 1.1).
- ModelSamplingAuraFlow shift: Adjusts sigma schedule (default 3).
- KSampler steps: More steps = higher quality but slower.
Execute
- Click Queue Prompt.
- One image will be saved.