Профиль

Not really. Answering webglaze requests in between other work. You can send dms and I’ll get to them when we can.

What Maya said. It is no easy task to maintain a tool built out of research, and maintain it over the course of years. So it is not surprising that there are few choices out there. Very few people understand the research, and even fewer have the $ to burn to maintain a tool that is free to all users

we answer all the requests we can when we can. Sometimes the social platforms delete requests that we don't get to quickly enough, and we lose them. So honestly, I don't know how to answer that q. At the moment, we're caught up on IG, but not on Bsky.

Surprisingly I think it was recorded, and the recording should go online at some point!

Update. The panel went exceptionally well. Rob was amazing as a moderator and we covered a lot of ground and topics in 50 mins. The room was completely packed, completely engaged, and we’re hoping for a bigger ballroom next year!

We ourselves do not have capability to hack hardware drivers. So we’re stuck waiting for PyTorch to update their libraries. We’ve looked into alternatives (unsupportive drivers) in the past but they have been brittle and problematic.

I’ve commented on both issues at various places but easy to miss. Linux is hard bec it requires sharing source code that makes tools easier to reverse engineer. Also supporting compatibility on a range of distros is something I’ve tried in the past and failed. We need PyTorch support for AMD. :(

Quick note. One of the earlier posts on social media had the wrong time for this event. It’s at 2pm local time. Hope to see you there!

So much of our work at ETCH/SAND Lab focuses on helping working creatives from AI harms. Dana's work focuses on an even more vulnerable population, children. I'm so glad Dana has written this book. It is a must read for every parent today. Just read the prologue, and you'll understand what I mean.

I've been reading Dana's new book on how to teach and nurture children in a world inundated by AI. It is such a powerful message that needs to be heard by every parent, communicated clearly, intuitively and emotionally. humanraised.org

Human Raised: Nurturing Connection, Curiosity and Lifelong Learning in the Age of AIHuman Raised is an essential guide to parenting in the age of AI from NYT best-selling author and pediatric surgeon Dr. Dana Suskind.humanraised.org

What exactly are the new models they’re referring to? Newer architectures are generally in commercially hosted models and cannot be used to do style mimicry. But this is not new. People like to generally claim Glaze doesn’t work any more, without having much of any clue about model architectures.

Sorry. We had some department email server issues. That’s been fixed and your results should be in you inbox now. They were queued up while the server got stuck.

Title aside, the panelists are all advocates for human creativity over AI.

Mist is very similar to Glaze. Where Glaze projects the image to a different artistic style, Mist projects the image to all black (to the best of my knowledge the last time I checked). It can be more effective on some AI operations (img2img). In studies that compared the two, results vary.

Ok, good, you know about watermarks. Didn't mean to jump in randomly on your thread :-)

just FYI, watermarks are useless against AI, and easily removed (google "image watermark remover"). Glaze & Nightshade do different things. Glaze prevents someone from fine-tuning/training a model to explicitly copy your style. Nightshade is a deterrent that disrupts models that train on your work.

Thanks Brian. For some reason we’ve never seen that bug before. Sounds like a goofy but easy to fix bug. Thanks!!

Thank you so very much! We are not yet able to accept donations online. We should be able to do this in 1-2 months. Thank you for your continued support!!

Check dms

We test whole songs. Samples will not produce false positives. That’s not how the detector works. And there’s no partial %. Detectors that give you those are giving you confidence values, not partial (35% of this was AI)