When publishing your artwork online, you risk it being used without permission. A watermark is a common solution to this problem.

By adding a pattern, signature, or text such as "Sample" as a watermark, you can make it harder to for others to use your artwork. These strategies can also have notable effects on images made by AI image generators that have been fine-tuned with said images, causing the watermark pattern to appear as an artifact, or reducing the accuracy of the imitation.

Adding a watermark to your work

In Clip Studio Paint, when you choose Export (Single Layer) or Export timelapse, you can add a watermark to your artwork and export it in a format ready for social media.

1. Select the Export menu

To export images, select the desired file format from the File Menu > Export (Single Layer). After you specify the save destination, the Export Settings dialog will appear.

To export a Timelapse, select the File Menu > Timelapse > Export timelapse.

2. Select the watermark pattern

At the bottom of the Export settings dialog, turn on ① Watermark and/or ② Noise pattern, then tap Watermark settings.

① Watermark: You can apply any image as a watermark.

② Noise Pattern (Ver.3.1 and later): You can apply a noise pattern to your artwork.

3. Set a watermark

In the Watermark Settings dialog box, you can adjust the intensity of the Watermark and Noise Pattern effects.

Watermark

Import any image from your Files or Photo Library and specify the position, size, and angle using the handles on the preview screen. You can also adjust the opacity and blending mode.

Noise pattern

Increasing the Strength will make the noise more prominent. You can adjust the value with a slider while previewing the effect.

Once you have decided on the settings, tap OK to close the Watermark settings dialog.

4. Export Images and Timelapse

Tap OK in the Export Settings dialog to export.

Once the Export preview dialog box appears, check the image and tap OK.

Your settings are saved in the app, so the same settings will be applied when exporting other projects.

Countermeasure against generative AI model fine-tuning

Watermarks can be an effective countermeasure against fine-tuned generative AI models.

While results will vary based on the color, strength, and pattern of the watermark and the prompt used to generate the image, a watermark can cause the pattern to appear in the generated image or lower the accuracy of the style imitation.

 

What is fine-tuning?

 

Fine-tuning is the process of training an AI model on a dozen to few dozen images with a specific style or character, allowing it to more accurately generate images that mimic that specific style or character.

Let's test this using three different watermarks (A: Foliage scroll pattern, B: Stripes, C: the Clip Studio Paint logo) with the following settings.

Watermark Settings

Opacity: 20%

Blending mode: Normal

Tiling: Repeat

Tiling direction: Vertical and horizontal

 

Illustrations used to check the effects

We prepared 15 illustrations by the same artist to fine-tune an AI image generator. These illustrations were newly created in-house and were used exclusively for this purpose.

Using the AI image generator, we generated an image using the prompt "sks style, 1girl, brown hair, masterpiece, best quality, portrait", which resulted in the following images:

 

Results by watermark pattern

 

None

 

A

 

B

 

C

 

While the strength depends on the watermark, we can see that the watermarks appear in the generated images.

 

Taking measures using noise patterns

When works with a Noise pattern are used for fine-tuning, the quality of the generated images can be expected to decrease. As a rule of thumb, the stronger the noise, the more effective it is at countering fine-tuning.

 

Using the same illustrations as before, we added various levels of noise and fine-tuned a model for each intensity.

 

Noise pattern settings

Noise intensity: None, 20, 40, 60

 

When the AI image generator was prompted to generate the image using the same terms as before, the following images were generated.

 

Results by noise intensity

 

None

 

20

 

40

 

60

 

While the style was imitated with None, we can see as we increase the intensity of the noise, the fidelity, quality, and even composition accuracy of the generated images decreases.

 

The effectiveness of watermarks and noise patterns will vary depending on the AI model and fine-tuning algorithms use to generate images. As such, we cannot guarentee that the expected results will be achived in all situations.