Remove or refill parts of an image with a small open model

Moebius is an open image inpainting model. Inpainting means you keep most of an image, mask the area you want changed, and ask the model to generate only the missing or unwanted region.

That makes it different from a generic prompt-only image generator. If you have a room photo with an object to remove, a portrait with a damaged background, or a scene where one area needs repair, Moebius gives you a focused edit instead of forcing you to regenerate the whole image. The public Space includes variants for natural scenes and faces, so the first test can match the kind of image you care about.

Start with the public Moebius Inpainting Space. Upload a clear image, paint over one region, and choose the closest variant: Places2 for natural scenes, CelebA-HQ or FFHQ for faces, or Pretrained for a more general test. For a useful first run, mask a medium-sized object or background gap and check whether the fill matches the surrounding texture, lighting, and perspective.

For local use, the backing model weights, Space code, and GitHub project are public. The workflow loads Moebius weights with a PixelHacker VAE, so it is reproducible from open components. The practical limit is hardware: local testing still expects CUDA, which makes the browser demo the right first stop.

The main failure mode is local mismatch. Watch for fills that blur texture, bend perspective, or change identity around faces. A small mask on a simple background is a fair first test; complex objects and large missing regions are harder.

• Remove distracting objects from product, room, or campaign images.
• Repair missing or damaged regions in a photo without rebuilding the whole scene.
• Compare face-focused and scene-focused inpainting variants on the same mask.
• Test whether a small open inpainting model is good enough before setting up a heavier image-editing workflow.
• Build quick before-and-after examples for visual cleanup tasks.

Try the demo

View model page