Watermark removal
Paint over the watermark · background rebuilt from neighbours · export at original resolution
点击选择或拖拽一张图片到此处(一次一张)
支持 JPG / PNG / WebP / GIF(首帧)/ BMP;单张不超过 20MB、2000 万像素;加载时按 EXIF 方向归一化
还有别处水印:直接在图上继续标记,再点一次「去除标记区域」。看着不满意:按下面的「能力边界」判断该不该换办法,或者直接「还原原图」重来。 满意了就点顶部的「下载 PNG / JPG」——按原图分辨率导出,只改标记过的地方。
能力边界:先看这一节,再决定要不要用
本站这个工具用的是浏览器本地的补图算法(多尺度加权扩散):它只会拿标记区周围的像素往里面外推, 不会凭“想象”画出图片里本来就没有的东西。所以效果好坏几乎完全取决于水印压在什么样的背景上。下面这三档是实测结果,不是免责套话。
天空、纯色底、渐变、白底商品图、虚化人像、皮肤、失焦远景。这类背景本身没有细节,外推出来的结果与原位几乎一致,正常观看距离看不出补过。
墙面、布料、木纹、有柔和光影的物体表面。水印会消失,但那一小块会比周围更“平”,放大看有一片轻微的涂抹感;缩小到网页尺寸通常可以接受。
草地、树叶、碎石、密集织物、人群、已经压在一排文字上的水印。这些区域的细节无法从周围推算,任何本地算法补出来都是明显的一团糊。这种情况下更实在的办法是:用「图片裁剪」把带水印的整条边裁掉,或者用「图片打码」老老实实盖住它。
另两点说清楚:① 我们没有接 AI 修复模型(开源里最小的 MI-GAN 也要 27MB 权重,外加约 6MB 的 onnxruntime-web 推理引擎,手机上首次使用还要现下载,与“打开就用、文件不出本机”冲突),所以补不出纹理属于设计取舍; ② 处理是按标记的连通块一块一块算的,整张图铺满小字水印也能处理,只是要分几次标记。
FAQ
Neither. The page downloads no model and makes no network request related to your image: loading, marking, inpainting and exporting all run in your browser's memory via canvas - the server has neither your image nor any queryable processing record. Closing or refreshing the page erases the image from memory, so download before you do that.
Look at two sides. The watermark itself: semi-transparent text and logos, pure white or solid-color corner marks, and small text tiled across the image all work - it is enough that the mark covers the pixels. The background: over sky, solid colors, gradients, white backgrounds or blurred portraits the repair is barely visible; over grass, leaves, dense fabric or crowds (high-frequency texture) it leaves visible smearing - see the capability limits above.
It can be processed, but doing it in several passes is recommended. The tool splits your marked area into connected components and inpaints each independently: a few dozen small tiled marks take about one or two seconds on a desktop browser, longer on a phone. If one pass covers too much, the page tells you to split it up instead of making you wait on a spinning progress bar.
No. Writing happens only inside your marked area (plus the few pixels of 'edge expansion' around it); every byte outside stays exactly as it was - verified pixel by pixel by tests before release. The export is a newly encoded image at the original resolution; the file on your computer is never modified.
Not recommended - please do not do this. A watermark is usually precisely how a rights holder asserts rights over a work; removing it from someone else's work and redistributing may infringe copyright and the right of network dissemination. This tool is only for scenarios where you have the right to process the image: recovering the clean version of your own original, licensed stock you bought with permission, or your own product photos. Legality of use is the user's own responsibility.
They solve different problems: watermark removal tries to restore the background so nothing shows it was there; mosaic covers the content and admits it is blurred - right for hiding private information; cropping just shrinks the frame and cuts off the watermarked strip - the cleanest and fastest when the size allows. If the watermark sits against the edge of the image, try cropping first.
One image up to 20MB and 20 megapixels, the same limits as the other image tools on the site. Beyond that the page asks you to run image compression or batch resize first. Also, an excessive total amount of marking (half the image painted at once, say) is refused with a hint to split it into passes - a browser gives one page limited memory, and pushing it crashes the whole tab.
Yes: painting uses pointer events, so you can drag a finger over the watermark directly, and the export is at full original resolution too. Time depends on the marked area, not the total image size - a corner mark of a few tens of pixels takes tens of milliseconds on desktop, a full-image tiled watermark about one or two seconds, and a phone is roughly three to five times slower. While waiting, the status line first says it is reading pixels, then that it is inpainting.