Z-Image - Image Editing with Mask and LoRA Library
How to use
- Upload an image.
- Brush over the area you want to change. No mask = the whole image is redrawn.
- Write a prompt and press Generate.
Masking
Cover the whole object plus 5-10 px beyond its edge. If the replacement has a different shape (t-shirt to coat), also brush the area the new shape will occupy - the model only paints inside the mask.
Prompt
Describe only what goes inside the mask, not the person or the scene.
a black wool double-breasted coat, open collar, natural folds, soft studio light
a light blue linen shirt, rolled sleeves, relaxed fit, natural fabric texture
Press + match original lighting to append a lighting-consistency phrase. Without it, replaced objects usually look pasted on.
Strength
| Situation | Value |
|---|---|
| With a mask | 0.85 - 1.0 |
| No mask, restyle only | 0.2 - 0.35 |
| No mask, style change | 0.4 - 0.55 |
| No mask, heavy rework | 0.6 - 0.75 |
Model
Turbo is fast - 8 steps, guidance 0. Base is slower but handles fabrics and materials better - 30 steps, guidance 4.
Steps and guidance switch automatically. Selecting a LoRA also switches to the model it was trained for.
Troubleshooting
| Problem | Fix |
|---|---|
| Old colour shows through | Raise Strength to 1.0 |
| Visible seams | Widen the mask |
| Wrong object generated | Switch to Base |
| Bad anatomy at the edges | Generate again with a new seed |
LoRA library
Paste a Civitai model link and press Add LoRA. Name, trigger words and base model are read automatically and stored on persistent storage, so they survive restarts. Trigger words are prepended to every prompt; typical weight is 0.6 - 0.9.
Only LoRAs based on Z-Image (ZImageTurbo / ZImageBase) will work - anything else is
flagged with a warning. For paid or gated downloads, add a Space secret named CIVITAI_TOKEN.
Do not use this Space to create intimate or sexual imagery of real people without their consent.