Blog

Why GemClean? A side-by-side test on a public Gemini image dataset

Getting rid of the bright sparkle is only half the job. We looked at what four tools left behind in the same corner of each JPEG.

What we found

GemClean scored higher than all three other tools on 12 of 20 images. It scored higher on 13 against GeminiWatermark.io and on 13 against GeminiWatermarkRemover.io. Gradually AI returned no result for 3 inputs.

GeminiWatermarkRemover.io scored below 3 on 16 images. GemClean scored below 3 on 6. A score of 3 means the crop showed no readily visible residue or new artifact.

The difference was easy to spot on a blue-gray reflection. On an image with two overlapping sparkles, all four tools left a visible mark.

The team behind GemClean ran this comparison.

A sparkle over a reflection

The three browser tools leave a recognizable outline on the water. GemClean leaves a much fainter trace.

Original crop for a sparkle over a reflection
Original
GemClean crop for a sparkle over a reflection
GemClean
GeminiWatermark.io crop for a sparkle over a reflection
GeminiWatermark.io
GeminiWatermarkRemover.io crop for a sparkle over a reflection
GeminiWatermarkRemover.io
Gradually AI crop for a sparkle over a reflection
Gradually AI

Matched 256 × 256 crops. Original from Goku OpenLab.

See the tied result and the overlapping sparkles

A textured wall

All four results earned the same score. Some leave a faint outline in the stone, while GemClean slightly softens the texture.

Original crop for a textured wall
Original
GemClean crop for a textured wall
GemClean
GeminiWatermark.io crop for a textured wall
GeminiWatermark.io
GeminiWatermarkRemover.io crop for a textured wall
GeminiWatermarkRemover.io
Gradually AI crop for a textured wall
Gradually AI

Matched 256 × 256 crops. Original from Goku OpenLab.

Two overlapping sparkles

Every tool leaves a visible mark. GemClean reduces the larger shape, while the second sparkle remains.

Original crop for two overlapping sparkles
Original
GemClean crop for two overlapping sparkles
GemClean
GeminiWatermark.io crop for two overlapping sparkles
GeminiWatermark.io
GeminiWatermarkRemover.io crop for two overlapping sparkles
GeminiWatermarkRemover.io
Gradually AI crop for two overlapping sparkles
Gradually AI

Matched 256 × 256 crops. Original from Goku OpenLab.

How we ran the test

We drew candidates in a fixed random order from three folders of the Goku OpenLab Nano Banana Pro dataset. We selected 20 JPEGs with visible corner sparkles by inspecting the inputs. The same files went to three standalone tools we found in a desktop Google search for gemini watermark remover. Each site describes its method as reverse alpha blending. We used each site’s standard workflow.

We ran the same P20 model used by GemClean locally in FP32 at the service’s 256 × 256 crop size. Codex scored matching crops, with tool names hidden and the order shuffled for each image. Some images had appeared in earlier testing. The 0–3 scale assessed visible residue and changes to nearby detail.

A tie means two outputs fell in the same grade. We used files as published and could not verify their export history or whether they were disjoint from training data. The scores describe these 20 JPEGs from one collection.

See all 20 scores, the scoring scale, and downloads
Compared withGemClean higherSame scoreGemClean lower
GeminiWatermark.io1370
GeminiWatermarkRemover.io1370
Gradually AI1820

Scoring scale

  • 3 — No readily visible residue or new artifact
  • 2 — Faint residue or small local change
  • 1 — Obvious residue or new artifact
  • 0 — No result or major visible defect
Source imageGemCleanGeminiWatermark.ioGeminiWatermarkRemover.ioGradually AI
NB_10220_03111
NB_10369_01000
NB_10081_03222
NB_10049_03331
NB_10020_03130 No result
NB_10338_03221
NB_15562_03111
NB_15807_02220 No result
NB_15188_03111
NB_15773_02200
NB_15003_03111
NB_20273_02000 No result
NB_20632_02222
NB_20450_03112
NB_20528_03333
NB_20869_03222
NB_15386_03111
NB_10458_02220
NB_15288_03332
NB_10413_03011

GemClean checkpoint pass020-980caaf7-fp32-v1. Model weights are private. The JSON includes SHA-256 hashes and links to the pinned dataset revision.

Download all comparison crops (ZIP)Download scores and input hashes (JSON)

Inputs by Goku OpenLab under CC BY 4.0. The figures show cropped inputs and tool outputs.

Try GemClean on your image