Type it. Select it. AI paints it in.

How generative fill works here

The editor runs SD-Turbo, a fast distilled Stable Diffusion model from Stability AI, directly in your browser with WebGPU. When you generate, the area around your selection is taken as context, the selection is pre-filled from its surroundings, and the model then denoises it in a few steps guided by your prompt — while everything outside the selection is re-imposed from your photo at every step. Finally, the colour along the edge is matched so the new content meets the photo without a seam.

The model (about 2.6 GB) downloads once and is kept by your browser, so later sessions start quickly and work offline. Your photo and prompt never leave your device.

What you can do with it

• Add objects: a boat on a lake, a bench in a park, birds in the sky, a lamp on a desk.
• Replace things: a plain wall with bookshelves, an empty table with a bowl of fruit.
• Restyle a region: raise Keep original shapes to change a jacket's material or a car's colour while keeping its shape.
• Remove things: leave the prompt empty and the selection is filled from the surroundings.

Tips for realistic results

• Select a little more than you need with a loose lasso; hard rectangular edges are easiest to spot.
• Describe only what goes in the selection: “a small white sailboat, side view”.
• Generate a few variations — each uses a new random seed — and keep the best.
• The result is a layer, so you can lower its opacity, mask its edges or clip an adjustment to it.
• Read the full generative fill guide for prompt examples.