[Bug]: Shutterstock watermark everywhere since version 1.2.0, but absent in 1.1.0
Hellisotherpeople opened this issue · 12 comments
Is there an existing issue for this?
- I have searched the existing issues and checked the recent builds/commits of both this extension and the webui
Have you read FAQ on README?
- I have updated WebUI and this extension to the latest version
What happened?
Images generated with version 1.2.0 full of shutterstock watermarks, this literally never happened with version 1.1.0
Steps to reproduce the problem
Update to version 1.2.0
do literally anything
What should have happened?
Be like version 1.1.0 but with fixed bugs
Commit where the problem happens
webui: Latest
extension: 1.2.0
What browsers do you use to access the UI ?
Mozilla Firefox
Command Line Arguments
unimportantConsole logs
unimportantAdditional information
You claim you "fixed the injection" in the readme but if this is fixed than I want it broken again.
Why did you close this? version 1.1.0 is literally better in everyway. It doesn't matter if it's "bugged", I want the "bugged" version.
The collab notebook also doesn't have this problem...
I will likely fork your 1.1.0 codebase and simply fix it as it breaks due to automatic1111 extension updates if you're going to let your own extension be like this.
Please visit the link I sent you. This is the problem of the training data, which means that it is not something that I am able to resolve. v1.1.0 is not a correct implementation. I had inserted motion modules to wrong place due to a typo.
If you generate GIF using the original repo, you will also see watermarks for at least some samples because I see them too.
If you like the previous version, I can create another branch which reserve the accidental “wrong” implementation. However, I’m not sure how to actively maintain both branches without double work.
@Hellisotherpeople would you be able to generate an example with identical settings between both versions?
@continue-revolution if the older version does operate better, and unless someone takes the time and resources to retrain the motion model on a non-watermarked dataset, the "proper" implementation is nearly unusable.
We may however be able to figure out exactly why your previous method worked better, and purposefully attempt to capture only the motion from their trained model and not the static watermark.
@AI-Casanova because I inserted motion modules into the incorrect place due to a typo (wrote a 2 where it should be a 3). Again, in v1.2.1 I will give you the option to choose whichever you like. If you like the previous one, it should be trivial to switch back on your webui.
@Hellisotherpeople is there a new fork you're working on? Can to tell me the best commit to pull before the watermarks start to show up?
SO basically this extension is useless because the creator refuses to remove the watermark. Most likely because they plan to monetize it. Got it.
SO basically this extension is useless because the creator refuses to remove the watermark. Most likely because they plan to monetize it. Got it.
this extension creator has nothing to do with it, the model creators of animatediff trained on Shutterstock data hence the watermark
Now that some motion modules do not create this water mark, also the official v2 motion module do not create this watermark as well. Will close this issue for now.