OS Insert

OSInsertModel is an object insertion model that inserts a foreground object into a background image under a given bounding box.

It supports two inference modes:

• conservative: builds bbox mask directly on background, then runs insertion; faster and more stable.

• aggressive: uses ObjectStitch + SAM + insertion refinement; better flexibility for hard cases.

API

from libcom import OSInsertModel

model = OSInsertModel(device='cuda:0')
result = model(
    background_path='path/to/background.png',
    foreground_path='path/to/foreground.png',
    foreground_mask_path='path/to/foreground_mask.png',
    bbox=[1000, 895, 1480, 1355],
    result_dir='path/to/save_dir',
    mode='aggressive',  # 'conservative' or 'aggressive'
    verbose=True,
)

Key Arguments

• background_path: background image path.

• foreground_path: foreground image path.

• foreground_mask_path: foreground mask path.

• bbox: insertion box in [x1, y1, x2, y2] format.

• result_dir: output directory.

• mode: 'conservative' or 'aggressive'.

• seed: random seed (default 123).

• strength: insertion strength (default 1.0).

• split_ratio: denoising split ratio (default 0.5).

• verbose: whether to save intermediates into result_dir/intermediates.

Output

Returns composed image as numpy.ndarray (or None when inference fails).