给自己生成的 NotebookLM / Gemini Notebook 导出文件去可见水印:图片、Gemini 图、PDF、PPTX、音频在本页内处理,文件不离开这台设备,全部处理代码都在页内、断网可用;视频生成 FFmpeg 命令在本机终端跑,同样不上传。
Remove the visible watermark from NotebookLM / Gemini Notebook exports you created yourself: images, Gemini images, PDF, PPTX and audio are processed on this page — files never leave this device, and all processing code lives in the page itself, so it works offline. Video generates an FFmpeg command to run in your local terminal, also with no upload.
这一路处理普通图片与信息图右下角的水印;Gemini 原图的四角星请切到「Gemini 图」标签页,那边有固定尺寸与自建模板恢复。
This tab handles the bottom-right watermark on regular images and infographics. For the Gemini sparkle, switch to the “Gemini Image” tab — it has the fixed-size area and bring-your-own-template recovery.
原理:扫描右下角区域,以中位灰度为背景找偏离块,按连通域过滤掉噪点(面积 <20px)、细线(高 <8px)、边框线(长宽比 >12:1)与扫描框顶部四分之一内的块,同行块合并后外扩,再用矩形上下各 8px 条带的列平均做逐列垂直渐变填充。平面或渐变背景接近隐形;照片、纹理、图表线条下方会留一块平滑色块,属预期。
How it works: scan the bottom-right area, find blocks that deviate from the median-gray background, filter connected components — dropping noise (area <20px), thin lines (height <8px), border lines (aspect ratio >12:1) and anything in the top quarter of the scan box — merge same-row blocks, expand, then fill with a per-column vertical gradient averaged from 8px strips above and below. On flat or gradient backgrounds it is nearly invisible; over photos, textures or chart lines it will leave a smooth patch — that is expected.
四角星固定在右下角:短边 ≤1024 时方块 48px、边距 32px;短边 >1024 时方块 96px、边距 64px——固定两档,不随图片尺寸缩放(v1 曾按比例放大框,大图会吃掉正常画面,v2 已修)。先在上方拖入一张 Gemini 导出的图片,再选清理方式。
The sparkle sits at a fixed spot in the bottom-right: a 48px square with a 32px margin when the short side is ≤1024, and 96px / 64px above that — two fixed sizes, never scaled with the image (v1 scaled the box up and ate into large images; v2 fixes that). Drop a Gemini export above first, then choose a method.
模板只在内存里用、不上传、不保存。页内从模板提取 alpha(带透明通道就用透明通道,否则按亮度推),再按 original=(watermarked−α×255)/(1−α) 逐像素反推。注意误差放大:α=0.8 时约 5×、α=0.9 时约 10×,模板越准、恢复越干净;模板不准时角落会留噪点或淡方框,改用检测填充更稳。
The template is used in memory only — never uploaded or stored. The page extracts alpha from it (the alpha channel when present, otherwise luminance) and reverses each pixel with original=(watermarked−α×255)/(1−α). Beware error amplification: about 5× at α=0.8 and 10× at α=0.9, so the better the template, the cleaner the recovery; with a poor template the corner keeps noise or a faint box — detect-and-fill is the safer fallback.
页内代码直接解析 PDF 里的整页图片(NotebookLM 导出的 PDF 每页就是一张扁平图片),逐页用同一检测器清理右下角,再逐页重建成图像型 PDF:版面保留,但文字不可选、不可搜索、链接失效。只适用于这种图片页 PDF;带真实文字层的复杂 PDF 请用桌面工具。能选 PPTX 导出就优先选 PPTX。
In-page code parses the full-page images inside the PDF (NotebookLM PDF exports are exactly that — one flat image per page), cleans each page's bottom-right corner with the same detector, and rebuilds an image PDF page by page: layout is preserved, but text is not selectable or searchable and links stop working. This only suits image-page PDFs; for complex PDFs with a real text layer, use a desktop tool. If you can export PPTX instead, prefer PPTX.
动手前先做两秒测试:在 PowerPoint 里点一下徽标——能单独选中带手柄,就直接按 Delete,不用这个工具;点中选中的是整张背景图,才是烤进图片、走这里。母版里的徽标去「视图 → 幻灯片母版」里查。
Do the two-second test first: click the badge in PowerPoint — if it selects on its own with handles, just press Delete, no tool needed. If clicking selects the whole background image, the badge is baked into the image and belongs here. For badges in the master, check View → Slide Master.
用页内自带的 ZIP 读写在内存里打开,只处理 ppt/media/ 下检测到水印的图片并同名写回;幻灯片 XML、动画、备注、母版等其余条目原样通过。已知问题:WebP 图片会以 PNG 数据写回但保留 .webp 文件名,个别 PowerPoint 版本可能提示修复。
Opens the file in memory with the page's built-in ZIP reader/writer, processes only images under ppt/media/ where a watermark is detected, and writes them back under the same names; slide XML, animations, notes, masters and everything else pass through unchanged. Known issue: WebP images are written back as PNG data but keep the .webp filename, and some PowerPoint versions may prompt a repair.
NotebookLM 音频的水印是末尾的口播免责声明(约 3–5 秒),不是开头。先听最后 10 秒再定秒数:免责声明长于剪除秒数会残留,尾音衔接紧会误剪正文最后一句。
The NotebookLM audio watermark is the spoken disclaimer at the end (about 3–5 seconds), not the start. Listen to the last 10 seconds before choosing: trimming too little leaves part of the disclaimer; trimming too much can cut the last sentence if the ending is tight.
原文件试听(先听末尾再定秒数):Original (hear the ending before choosing seconds):
剪切后波形与试听(当场验货):Trimmed waveform & playback (check it right here):
浏览器版用 Web Audio 解码后导出 WAV(无损,但文件会比 MP3 大,约 10 MB/分钟;多声道源会混为双声道导出,处理前会先提示)。想要原格式流复制零损失,用下面这条命令在本机跑:
The browser version decodes with Web Audio and exports WAV (lossless, but larger than MP3 — about 10 MB/min; multi-channel sources are mixed down to stereo, with a warning before processing). For a lossless stream copy in the original format, run this command locally:
ffmpeg -i input.mp3 -ss 0 -t [duration-5] -c copy output.mp3
视频走 FFmpeg 的 delogo 固定坐标 + 剪片尾,需要全片重编码,浏览器内跑 WASM 版对内存和时长都很吃力,所以这里生成命令、你在本机终端跑——同样不上传。坐标只有两档、按帧高查表,不做检测;竖屏坐标原站未公布,不要按横屏比例推算。
Video uses FFmpeg delogo at fixed coordinates plus an end trim, and needs a full re-encode. Running the WASM build in the browser is heavy on memory and duration, so this page generates the command for you to run in a local terminal — still no upload. There are only two coordinate presets, looked up by frame height, with no detection; vertical-video coordinates were never published by the original site, so do not scale them from the landscape values.
-t;文件名里的引号、$、反引号已做转义处理。-t that would fail. Quotes, $ and backticks in filenames are escaped.参数来源:Skill muzi-notebooklm-cleaner,2026-11-22 实查;与页内实现逐项对照一致。
Source of parameters: the muzi-notebooklm-cleaner Skill, verified 2026-11-22; each row matches the in-page implementation.
| 格式Format | 关键参数Key parameters |
|---|---|
| 视频 1080pVideo 1080p | delogo x=1104, y=656, w=770, h=62; 剪尾 2.5s;CRF 18trim end 2.5s; CRF 18 |
| 视频 720pVideo 720p | delogo x=736, y=437, w=513, h=41 |
| 页内解析整页图片(DCT 直接取 JPEG、Flate 原始像素重绘);扫描框最小 350×80,大图至宽 22%·高 8%;阈值 80/60/45;连通域过滤(面积 ≥20、高 ≥8、长宽比 ≤12:1、排除顶部四分之一);外扩 +15/+8(PDF 标签页内可调);95% JPEG 逐页重建In-page parsing of full-page images (JPEG taken directly for DCT, raw pixels redrawn for Flate); scan box min 350×80, up to 22% W × 8% H on large images; thresholds 80/60/45; connected-component filtering (area ≥20, height ≥8, aspect ≤12:1, top quarter excluded); expansion +15/+8 (adjustable in the PDF tab); rebuilt page by page as 95% JPEG | |
| PPTX | 页内 ZIP 读写只改 ppt/media/;PNG 无损、JPEG 0.95 写回Built-in ZIP read/write changes only ppt/media/; PNG lossless, JPEG written back at 0.95 |
| 图片检测Image detection | 与 PDF 同一检测器;外扩默认 15px(图片标签页内可调);填充取上下各 8px 条带列平均渐变Same detector as PDF; padding defaults to 15px (adjustable in the Image tab); fill is a per-column gradient from 8px strips above and below |
| Gemini 图Gemini images | 固定方块 48px(短边≤1024,边距 32)/ 96px(边距 64),不缩放;自建模板 alpha 反推,误差放大 α=0.8 时 5×、α=0.9 时 10×Fixed square 48px (short side ≤1024, margin 32) / 96px (margin 64), never scaled; alpha reversal with your own template; error amplification 5× at α=0.8, 10× at α=0.9 |
| 音频Audio | 末尾默认剪 5s、开头 0s;导出 WAV(多声道混为双声道,先提示);流复制命令 -c copy;MP3 帧边界约 26msDefault trim: 5s off the end, 0s off the start; WAV export (multi-channel mixed to stereo, with a warning first); stream-copy command -c copy; MP3 frame boundary ≈26ms |
点一下,页内现造一张带模拟水印块的测试图(浅色渐变底 + 右下角灰字块,另在角落放一颗孤立噪点和一条细线验过滤规则),走一遍检测 + 填充,把前后对比和填充框尺寸摆出来。噪点与细线不该把填充框撑大。
One click generates a test image in-page (light gradient background, a gray “watermark” block in the bottom-right, plus an isolated noise dot and a thin line in the corner to exercise the filters), runs detect + fill, and shows the before/after with the filled box size. The noise dot and the thin line must not stretch the fill box.
原站 notebooklmremover.org 是闭源纯静态站(Astro + Cloudflare Pages),未找到仓库与许可证,只能对照功能、不能抄实现。本页的方法以 MIT 开源项目为准:
The original site, notebooklmremover.org, is a closed-source static site (Astro + Cloudflare Pages) with no repository or license found — features can be compared, implementation cannot be copied. This page's methods follow MIT open-source projects:
| 项目Project | 许可证License | 覆盖Coverage | 本页取了什么What this page took |
|---|---|---|---|
| alpinist-GH · Logo Remover v0.2.4 | MIT | 视频 + PDF + PPTX + 图片Video + PDF + PPTX + Images | 视频 delogo 快档思路、按格式分流的整体结构Fast-preset delogo approach for video; overall per-format routing structure |
| froggeric · wmr (132 星stars) | MIT | 视频 / Gemini 图Video / Gemini images | 反向 alpha 方法与「视频须 inpainting」的判断Reverse-alpha method and the “video needs inpainting” judgment |
| Albonire (83 星stars) | MIT | 批量 + 预览Batch + preview | PPTX 批量处理与 Telea 参数参考PPTX batch handling and Telea parameter reference |
| encoreshao · nlm-unwatermark | MIT | PPTX / PDF | 检测 → 干净 patch 优先 → inpainting 兜底的三段骨架Detect → clean patch first → inpainting fallback, three-stage skeleton |
| dovetaill / CourseBar101 | 无许可证No license | 图片Images | 只学双旋钮参数设计,不用代码Two-knob parameter design only; no code used |
| LazySo / ijpg / Source Helper | 未开源Not open source | 图片 / 桌面Images / Desktop | 只学方法,不用代码Methods only; no code used |
实现全在页内:图片走 Canvas API,PPTX 走页内 ZIP 读写,PDF 走页内解析与重建,音频走 Web Audio;只有视频 / 音频的原格式处理借你本机的 FFmpeg(命令由本页生成)。没有任何第三方库、没有 CDN 依赖。
Everything runs in-page: Canvas API for images, a built-in ZIP reader/writer for PPTX, built-in parsing and rebuilding for PDF, Web Audio for audio; only the original-format video/audio path borrows your local FFmpeg (with commands generated here). No third-party libraries, no CDN dependencies.