Manual Cleanup is the browser tool for the cases where a fixed detector is not enough. You open a saved image, paint over the area you want to fix, run cleanup and download the result. It is made for visible marks, small logos, captions, date stamps and small unwanted objects.
The important part is control. Instead of asking a generic online tool to guess where the watermark is, you show Banana Clean the exact area. That makes the workflow useful for shifted Gemini marks, busy textures and one-off images from other AI tools.
When to use Manual Cleanup
Use the automatic extension when you are downloading fresh images from Gemini or Google AI Studio. It is faster and tuned for the standard visible mark. Use Manual Cleanup when the image is already on your device or the automatic pass leaves something visible.
Good manual cleanup cases:
- Gemini watermark moved away from the usual corner
- Nano Banana mark is still faintly visible on marble, leather, fabric, sky or water
- A logo, caption, date stamp or small object needs to be removed from a saved image
- You downloaded a lower quality preview before installing the extension
- You want to inspect the before and after result before saving
It is also useful for edge cases we need to learn from. If you choose to share a sample, Banana Clean receives the original image and your painted mask, which helps improve future detection and cleanup. That upload is optional and separate from the local cleanup itself.
How it works for the user
- Open Manual Cleanup.
- Load a PNG, JPEG or WebP image.
- Brush over the watermark, logo, caption or object.
- Run cleanup in the browser.
- Compare the result and download the cleaned PNG.
You do not need to understand masks, alpha values or model internals. The brush is there so you can tell the cleaner what part of the image matters. The result depends on the background. Smooth walls, sky, fabric and simple gradients are usually easier. Tiny text, faces, hands, repeating patterns and sharp product edges are harder.
Why a manual brush beats guessing
Watermarks are not always fixed. Gemini images can place the visible mark slightly differently depending on aspect ratio, composition and export path. A pure corner-based cleaner can miss that. A generic AI watermark remover can over-edit highlights, reflections or small bright objects because it has to guess what is a watermark.
A brush removes the guess from the first step. You decide the region. The model only has to rebuild that area. That is safer for images with reflections, product shots, leather grain, bathroom tiles, foliage or dark backgrounds.
Privacy model
Manual Cleanup runs in the browser. The image is loaded into the page, the brush mask is created locally, and the cleanup result is produced locally. Banana Clean does not upload the image just because you used the tool.
There is one separate option: sample sharing. If you explicitly choose to help improve Banana Clean, the page can send the original image and mask as a product sample. That is for difficult cases where the cleaner needs better coverage. No checkbox, no sample upload.
What it does not do
Manual Cleanup removes visible content. It does not remove invisible AI provenance systems such as SynthID. It also does not guarantee a perfect result on every background. A visible watermark over detailed text, a human face or a sharp product logo can still need a second pass or a dedicated editor.
The tool is best when the target area is small and the surrounding image gives the model enough context to rebuild the patch naturally.
Where it fits inside Banana Clean
Banana Clean now has three practical paths. The extension cleans supported Gemini and AI Studio image downloads automatically. OMNI Video handles supported Gemini and Google Flow MP4 downloads as a separate plan. Manual Cleanup covers saved images, shifted marks and leftovers.
If you are starting with a fresh Gemini image, install the extension. If you already have the file, use Manual Cleanup. If you are working with video, read the Gemini and Google Flow video cleanup guide.