Inpainting mask and prompt guide
Inpainting replaces a masked region of an image with new content — fixing a hand, swapping an object, removing a distraction. The hard part isn’t the mask; it’s getting the new pixels to blend instead of looking pasted on. This guide turns your “what’s there / what I want” description into a tailored prompt and the right denoising strength.
What makes inpainting hard
The blend failure happens because the model generates the masked region with some independence from the original image. If your prompt doesn’t explicitly anchor the new content to the surrounding lighting, style, and color palette, the model satisfies the content description but ignores the context — and the result looks cut-and-pasted.
The second common failure is denoising strength. Too high, and the model regenerates so aggressively that it changes things you wanted to keep (surrounding texture, colors bleeding in). Too low, and the edit is too timid — the original content shows through or the change looks ghosted. Finding the right denoising range for your specific edit is the core skill in inpainting.
How the tool works
Two inputs govern a good inpaint: the prompt and the denoising strength. The prompt should describe what you want in the masked area plus enough of the surrounding scene — lighting direction and quality, dominant colors, artistic style — for the generated pixels to harmonise. Denoising strength controls how far from the original pixels the model can stray: low values (0.2–0.4) make subtle fixes; high values (0.7–1.0) fully replace the region.
This guide combines your description of what is currently there and what you want instead, then outputs a structured prompt that includes a seamless-blend request and suggests a denoising band appropriate to the size and nature of your change.
Denoising strength guide
| Change type | Suggested denoise range | Notes |
|---|---|---|
| Texture or color tweak | 0.2–0.35 | Stay close to original content |
| Remove a small object | 0.4–0.55 | Allow background infill |
| Swap a similar object | 0.5–0.65 | Content changes, style preserved |
| Replace a face or person | 0.65–0.80 | Significant change needed |
| Fully regenerate region | 0.75–1.0 | Treat mask as a new generation |
These are starting points — step up in increments of 0.1 rather than jumping to 1.0.
Tips for seamless blends
- Match the light explicitly. Write “warm morning light from camera left, casting soft shadows” rather than just “sunlit” — specific lighting description pulls the blend tighter.
- Feather and oversize the mask. A slightly larger, soft-edged mask gives the model context and kills hard seams at the mask boundary.
- Start denoising lower than you think. The most common inpainting mistake is going too high too soon and over-cooking the whole masked area.
- Iterate the region only. Re-run just the mask with small prompt tweaks rather than regenerating the full image — it is faster and preserves the parts you are happy with.
- Use “original” or “fill” for masked content init. Most UIs offer latent noise vs original vs fill for how to initialise the masked area before diffusion. For blending, “original” or “fill” usually outperforms latent noise.