lthero-big
Does it handle image editing as robustly as VINE?
Thank you for your work on this interesting project!
I recently came across a watermarking model VINE, which emphasizes robustness to image editing (such as image inversion, text-driven editing, image-to-video generation, etc.).
I’m curious to understand if your method is robust against such kinds of editing.
If you’ve conducted any evaluations or comparisons under such conditions, I would really appreciate any pointers or insights.
I just generated some watermarked images by FLUX and edited by ledits and detected the bit accuracy, sadly, it just around 50%. I'm afraid it's not robust against these regenerated-based attacks.
There are additional details that I would like to share. After applying methods such as ledits to images generated by FLUX(flux1-dev), the watermark is undetectable( Bit Accuracy is around 50%). However, images generated by SDv2-1 maintain a high Bit Accuracy of approximately 99% after editing.
That's an interesting observation! Do you have any insights into why this might be happening?
Does this mean that this watermarking method is tailored for SD, and needs some improvement if wanna apply it to FLUX? Is there any way to combine VINE's robustness with this watermark method on FLUX?
ledits may be optimized for SD, allowing for better preservation of the latent space structure of SDv2-1 during the editing process, thereby come with high watermark Bit Accuracy. In contrast, editing with images generated by FLUX may significantly alter the latent space, reducing the watermark detection rate.
This makes sense. Thanks for your patient explanation! I will further explore it.