Removing watermarks and artifacts with inpainting
AI-generated and processed images frequently pick up unwanted marks: a residual watermark from a reference image, stray text the model hallucinated, or JPEG compression blocks from a low-quality source. Stable Diffusion’s inpainting mode is the most reliable way to clean these up, but the right settings depend heavily on what you are removing and on which model you run. This tool gives you a tailored configuration for each case.
How it works
You select the artifact type, the image type, and your available model. The tool then recommends a denoising strength, mask blur, only-masked padding, fill mode, and step count. Text and logo removal needs high denoising (0.55-0.7) so the region is genuinely regenerated; compression-artifact cleanup needs low denoising (around 0.3) so structure is preserved while blockiness is smoothed. Padding and mask blur are tuned so the reconstructed region blends seamlessly into its surroundings.
Settings guide by artifact type
Solid watermark or logo on a photo
The mark is visually distinct from the background and needs to be replaced with realistic image content. Use denoising strength in the 0.6-0.75 range, mask blur around 4 pixels, and only-masked padding of 32 pixels. The high denoise fully regenerates the region; the padding gives the model context about the surrounding texture. Run at 30-40 steps. If the logo sits on a busy background (foliage, fabric, complex textures), increase padding to 48 and check that the fill mode is set to “original” so the model has the real pixels to reconstruct from rather than a latent noise fill.
Stray text hallucinated by the model
AI image generators sometimes produce text-like shapes in signs, clothing, or background objects that are not coherent words. These differ from deliberate watermarks in that the surrounding content was meant to be there — you are patching a generation error rather than replacing a clear overlay. Use denoising in the 0.55-0.65 range, slightly lower than for hard watermarks, because the surrounding pixels were part of the intended image and a large shift in denoise will create a visible patch boundary. Mask tightly around the text shape with 2-pixel blur.
JPEG compression artifacts (blocking)
These require a completely different approach. High denoising changes the image content rather than cleaning the artifacts, making the result look worse. Use low denoising (0.25-0.35) paired with an upscaler at similarly low denoise (0.2-0.3). The goal is to smooth blockiness and restore detail without altering the underlying image. Never use high denoise for compression artifacts — you will regenerate the scene rather than clean it.
Gradient or semi-transparent watermarks
Semi-transparent marks (often used by stock image sites) are the hardest case because the content underneath is visible through them. High denoising removes the mark but also the detail under it. Use moderate denoising (0.5-0.6) and mask very precisely at the mark boundary. Multiple passes with a fresh mask often produce better results than one aggressive pass.
Mask strategy and tips
- Mask generously. Include a small margin around the artifact so the model has surrounding pixels to reconstruct a plausible fill.
- Use only-masked at full padding. This inpaints the cropped region at high resolution, which matters most for small corner watermarks.
- Iterate. A faint ghost after the first pass usually clears with a second pass on a fresh, slightly tighter mask.
- Different rules for compression. Do not crank denoising for JPEG blocks — low denoise plus an upscaler restores detail without changing the picture.
- Stay within your rights. Only edit images you own or are licensed to use.