 Hmm, here’s an idea: what if tools like Glaze and Nightshade were integrated into fediverse servers and/or clients and automatically applied whenever any images are posted… 🤔 Thoughts @dansup? https://nightshade.cs.uchicago.edu #ai #selfDefense #glaze #nightshade #fediverse #pixelfed #mastodon 28 comments
 @markhburton @ErikUden @aral yeah, and from the looks of the site, they're only distributing binaries too. Hard to integrate anything like this into a project like Mastodon without support for one or both of these.  @markhburton @ErikUden @aral that looks like a limitation which could be overcome, as the download link mentions pytorch, which is cross-platform @ErikUden @aral   From the description it looks like you'd need both, but the Nightshade people are working on a way to have only a single standalone tool that does both Still, depending on how heavy the tool is (not the slightest idea), it might actually increase instance costs, maybe not but would be a thing to check before integrating it with servers   @patterfloof @aral @dansup These tools are expensive to run in terms of CPU, it would be much better to convert the image once. @TomSwirly @aral @dansup yeah, I'm spitballing on the idea of arming every image on the internet, down to the UI icons on a website Instead of putting all the load on the instance, which is probably fairly heavily loaded, it would be a lot better if the upload tool did the conversion on the user's system, which almost certainly is idling nearly all the time.  @TomSwirly Oh, yes. But how many instances will have that, though? Most instances are run on shared virtual machines! My back-of-the-envelope 30-second calculations are that if an instance did this, it would cost between one and two orders of magnitude more CPU cycles than they are spending now and at least twice the RAM (and as video memory, as you point out). People aren't going to pay > five times as much to run an instance, I'm pretty sure. @TomSwirly @aral @dansup AFAIK this cannot be RAM, this needs to be memory on the actual video card. @TomSwirly @aral @dansup and expecting the client to do that is expecting people to go "oh, yeah, let me fire up my gaming desktop to upload the photos from" (assuming they didn't get an AMD card, in which case they need to ask someone else to upload photos for them).   @aral it took ~12 hrs for 1 images to process on the medium scale & close 7hrs for the default settings. it is awesome & i want to see more use…  @brentpruitt @aral I second this. I tried it on a 2019 MBP and it took 2-3 hours for a small jpg while taking all of my computing power. there also isn't any plan to port to linux which is a shame. I like the idea, but implementation is way off so far. Needs to be in the 10 minute/image range for me to consider using it on a regular basis imo. @aral anything to stop the theft! glaze has been making really quick version updates, so who knows… fingers crossed @aral I wouldn't count on it being intergrated into official releases of Fedi servers, as it doesn't seem to be an open source project. Maybe clients could if they choose to but then you won't find them for download in F-Droid. Individual instances could maybe integrate as well. I think it makes more sense for the user to process images themselves before upload if they wish.  |
Gregory's Server
@aral wow, first time hearing of Nightshade! This is an amazing tool...