Moonshine-Image Review: Mask-Free SLBR Removal for Semi-Transparent Watermarks

The Most Annoying Watermark Is the Semi-Transparent Kind

After trying enough watermark removal tools, you notice a pattern: solid marks — a plain corner badge, bold black text — are easy. Any AI tool cleans those up. The real headache is semi-transparent watermarks: white "PREVIEW" text, faint TV station logos, ghosted brand marks blended into the photo. They merge with the background, have no clean edge, and manual masking often goes wrong. AI inpainting frequently leaves a visible gradient trace where the mark used to be.

For a long time, this niche problem had no good answer. Then a Chinese open-source project called Moonshine-Image brought a 2021 academic algorithm back into the spotlight: SLBR semi-transparent watermark removal — no manual mask, one-click batch processing, fully local.

1. What Is SLBR: A Paper That Made Masking Optional

SLBR stands for Self-calibrated Localization and Background Refinement, from the ACM MM 2021 paper Visible Watermark Removal via Self-calibrated Localization and Background Refinement, open-sourced by the BCMI lab at Beijing Jiaotong University (bcmi/SLBR-Visible-Watermark-Removal, 254 stars on GitHub).

The core is a two-stage framework:

1. Self-calibrated localization: the first stage adaptively predicts the watermark mask and refines it through self-calibration — this is the "mask-free" magic, the model finds the watermark itself;
2. Background refinement: the second stage uses the mask to guide background enhancement, filling the watermarked region so it blends naturally with the surroundings.

In published comparisons (CLWD / LOGO30K benchmarks), SLBR reaches 38.28 / 41.50 dB PSNR and 0.9814 / 0.9914 SSIM, clearly ahead of contemporary methods like WDNet, BVMR, and SplitNet. For a 2021 method, those numbers still hold practical value — especially the no-manual-annotation property, which is exactly what batch workflows need.

2. Moonshine-Image: SLBR Packaged Into a Windows Desktop App

Moonshine-Image is a local Windows desktop tool for image/video watermark removal and object erasing, by developer CuiMuxuan. GPL-3.0, 135 stars, 79 commits, current version v1.2.0. Stack: Vue 3 + Quasar + Electron + a local Python backend — originally a secondary development of the well-known IOPaint project, now migrated to a built-in moonshine_server backend.

It ships three processing models covering different scenarios:

Model	Mask required?	Best for	Notes
Lama	Manual mask	General erasing, objects/text/irregular occlusions	Infers from surrounding content
MAT	Manual mask	Large-area repair, same tier as Lama	Needs CUDA; CC BY-NC 4.0 weights, non-commercial only
SLBR	No mask	Visible semi-transparent watermarks	One-click batch; poor on opaque marks

SLBR is the tool's biggest differentiator: drop a folder of semi-transparent watermarked images in, select SLBR, run, and batch output comes out — the brush never gets touched.

3. Capability Boundary: The Author States the Limits Upfront

The README opens with an important disclaimer: watermark removal is not original-image restoration. Watermarks obscure or corrupt original pixels; no removal model can fully recover what was underneath.

Concretely:

• SLBR only works on visible semi-transparent watermarks; it performs very poorly on opaque ones — for opaque stickers, bold black text, or logos covering key details, use Lama with a manual mask or generative-AI redraw;
• When watermarks cover faces, text, product details, the author recommends generative AI (Jimeng, Stable Diffusion, Flux) for redrawing;
• For video watermarks you can try ProPainter, but it's hardware-hungry and the author states it is not planned for integration.

This honest "know your limits" stance is refreshing — many tools overpromise and underdeliver. Moonshine-Image at least tells you exactly when to use it and when to switch tools.

4. Image Processing: Batch Is the Main Course

Image mode supports single image, selected files, folder mode, and large batch path import. Once loaded you can:

• Draw masks manually (brush / rectangle / eraser) for Lama-based irregular-region removal;
• Pick SLBR and run semi-transparent watermark removal with no mask at all;
• Compare before/after in real time;
• Choose output format: auto, original, png, jpg, webp — with quality control for JPG/WebP.

For e-commerce operators and asset managers dealing with hundreds of images, "folder batch import + mask-free SLBR" compresses hours of per-image work into a single run.

5. Video Processing: Timeline, Keyframes, Resume

Video cleanup is the other half of the tool:

• Visual timeline editing: import a video, set keyframes on the timeline, and control exactly when the mask applies — if the watermark only appears in the first 3 seconds, only that segment gets processed;
• Moving watermarks: the timeline lets you process segments separately as the mark shifts position;
• Resume after interruption: failed or interrupted runs reuse completed segments and continue instead of restarting (note: no multi-task queue);
• FFmpeg fallback export: if WebAV export fails, it automatically switches to FFmpeg so you still get a result.

6. Download & Install: Choosing Among Five Packages

Moonshine-Image ships Windows x64 builds only, split by Torch runtime and model strategy:

Package	Who it's for
cu130	New NVIDIA GPUs (especially RTX 50-series)
cu126	Older GPUs, or unstable CUDA 13.0 compatibility
cpu	No NVIDIA GPU; noticeably slower
bundled-models	Lama + SLBR included, fastest start, larger download
external-models	Smaller; models downloaded in-app on first run

New GPU? Pick Moonshine-Image-v1.2.0-win-x64-cu130-bundled-models.zip and verify with the SHA256SUMS.txt from the Release. Two download channels: Quark Pan and Hugging Face.

Install notes:

• Avoid non-ASCII paths: Chinese characters in the path can break Python or model loading — extract to a pure-English path;
• Missing models on first launch of an external-models build is normal — check "Global Settings > Model Management" for download sources;
• v1.2.0 is the last fully-offline release (bundles FFmpeg); future versions decouple the app, runtime, and model resources.

7. Who Should Use It

Moonshine-Image fits three profiles best:

1. Operators batch-cleaning semi-transparent watermarks — mask-free SLBR plus folder batch import turns hundreds of images into one run;
2. Privacy-conscious creators — everything processes locally, nothing is uploaded;
3. Light video users who want a free fix — timeline + keyframes handle most fixed or semi-moving marks, at zero cost.

If you have a stack of semi-transparent watermarked assets sitting around, skip the per-image upload dance at online tools and try this local option — because the only way to know if a watermark is truly gone is to zoom in on your own machine.