Single-Person Removal: 2-Pass Recovery Tutorial

Step 1: target definition

Goal: remove the blocked person in background and keep the centered woman in black.

Step 2: pass-1 mask (masking is correct)

The target person is fully marked; this masking step is correct.

Step 3: pass-1 failure diagnosis

Issue: both target and main subject were removed. This is model over-removal, not masking error.

Step 4: proper recovery (restart + prompt tuning)

Correct path: click restart to return to original image, then re-mask and add subject-preservation constraint.

Refined prompt example:

移除标记/着色区域中的人物。保留当前正中心的黑衣服女生，女生的人物手势姿势保持不动，用逻辑上延伸场景的智能背景填充替换它们。确保填充区域无缝连接——没有模糊的伪影、污迹或明显的编辑痕迹。保持与照片其余部分相同的颗粒感、景深和光照，以获得自然、未经编辑的外观。仅移除与标记/着色区域重叠的人物。未标记的人物必须保持不变。不要删除、替换或修改任何未标记人物。

Step 5: final pass-2 result

Result: centered woman is preserved, blocked person removed. Two passes delivered a stable outcome.

Key takeaways

• Pass-1 failure does not always mean wrong masking.
• When main subject is removed, restart first instead of refining on bad output.
• Add preservation constraints; do not rewrite default prompt completely.