lllyasviel
Regarding the Default Embedding of a Watermark in the Alpha Channel of Generated Images
With last week's commit, a feature was introduced that automatically embeds a watermark in the alpha channel of generated images. While this feature can currently be disabled through settings, I believe it is inappropriate for it to be enabled by default.
Some AI artists prefer to keep their prompts and techniques confidential, and there are users who do not want their alpha channels to be modified without their knowledge. Enabling the watermark by default could lead to unintentional exposure of information, which may constitute a privacy violation.
I strongly suggest that this feature be disabled by default, allowing users to manually enable it if needed. I would appreciate your consideration on this matter.
https://github.com/Panchovix/stable-diffusion-webui-reForge/issues/299
My opinion is this:
If a user wants their prompts confidential, they would have disabled embedding metadata to begin with, so as to not have anything potentially leak. Therefore, in the current implementation, alpha channel metadata is only included if infotext metadata is already enabled.
Websites do not scrub infotext metadata or alpha channel information consistently, so a user who truly wants their prompts kept hidden at all costs would have disabled metadata entirely already.
It can be assumed that a user who wants to keep metadata for themselves but scrub it before distribution elsewhere likely already has knowledge about all of this behavior, and can easily toggle these settings themselves, or use their own tools to prepare files for distribution.
keywords: stealth, privacy, invisible watermark, spyware, malware
This compromises the privacy of all Forge users who use metadata for their own purposes. There is an expectation that users can strip metadata for distribution, which is easy to do with Exif tags, and many sites strip them automatically. The same can't be said for hidden metadata like this, which is only a step away from an indelible watermark.
The PR for this was open for less than a minute before being merged. No discussion. No one requested it. The only "justification" is a link to an obsolete repo which was last updated two years ago and which clearly states that it's no longer needed for its original use case. The only purpose of this "feature" is to create a covert disclosure of the work's origin.
And now it has been landed without any users knowing. If there was a legitimate use for it, it should have been announced in the discussions and added to the news section. And, of course, it would have been opt in.
It's curious that it was added to Forge and ReForge but I can't find a corresponding PR for ComfyUI, despite the developer contributing to the latter...
This needs to be reverted, or, at best, made opt in, as has happened already in ReForge. @Panchovix?
@lllyasviel A few days ago, a NovelAI AI artist—believing he/she’d simply removed the metadata tags—didn’t realize he also needed to strip out the alpha channel. As a result, he was brutally slandered and bombarded with vicious attacks by the artist named in the tags and their 110,000-strong fanbase. This isn’t an isolated incident: because @catboxanon enabled an invisible watermark by default in Forge, every Forge user is now exposed to the exact same threat. If this critical flaw isn’t fixed immediately, Forge users will inevitably face the same ruthless persecution.