tkawano/object-removal

Object Removal (LaMa Inpainting)

Upload an image, paint over the object you want to remove, and LaMa fills the hole with plausible background — a content-aware erase. LaMa (Large Mask inpainting with Fast Fourier Convolutions) is resolution-robust and runs on CPU.

It's generative: it invents a believable background, not the true hidden pixels. Great for removing things (people, signs, blemishes); for text-guided replacement you'd use a diffusion inpainting model.

Run locally

pip install -r requirements.txt
python app.py

Open http://127.0.0.1:7860. On first use, the big-lama model downloads automatically.

How it works (short)

You give a 3-channel image + a 1-channel binary mask (white = remove). LaMa predicts the content under the mask in one forward pass. Fast Fourier Convolutions give each layer a global receptive field, so even large holes get a coherent fill — the key reason LaMa handles big masks better than classical patch methods.

The mask can come from anywhere

Here you brush the mask by hand, but any binary mask works — e.g. a Segment Anything cut-out (click to select an object, then inpaint its region). Select → remove.

Companion demos

• —click-to-segment — click any object (SAM) to get a mask
• —image-upscaler — general upscaling with Real-ESRGAN
• —face-restoration — restore low-quality faces (GFPGAN)

License

Apache-2.0 (app code). See the LaMa repository for model licenses.