Forum
>
Site Discussion
>
Thread
Log in |
Register
I think posts without names should not even be allowed on the site.
Not going to happen in the foreseeable future. In some cases the uploader might not know the name, in which case I'd still prefer to have the uploads than not, provided that they are good.
In other cases, uploaders have quirks, like tempp who for whatever reason doesn't tag them with a name, but his uploads are generally great and I much prefer the site with his 320K+ uploads than if they hadn't been allowed since he didn't tag them with a name.
First of all, that enables you to post girls that have requested not to be featured here.
Not really a concern, they usually quickly get tagged with a name by someone else which flags them for removal.
There is an another issue that works against the name argument: people finding named images that are fully AI or an AI version of a real person.
I’m not a fan of AI in most situations. At the same time I’m against using the hardline of “no name, no upload”, since sometimes a person genuinely doesn’t know the name and having a name is not a guarantee the image is legit anyhow. I’d prefer a name, but sometimes the community can help find an otherwise unknown name and quite honestly I’ve got better things to worry about than that.
The absent names can be an annoyance, but I've seen Chainer explain his reasoning before and agree. I've filled in a few names for Tempp unloads myself.
The AI slop has potential to be a problem, but AI also has potential to help counteract both the slop problem and the naming problem.
There are AI's specialized at identifying AI slop. There are also AI's specialized at facial recognition, naming, and finding.
But Chainer already has his own priorities and list of projects for GWM, and implementing an AI takes work.
tamarok★:
While there may be solutions to detect AI slop, I doubt they are free and are also likely computationally intensive. Maybe in the future, but now is unlikely the time.
You didn't follow my provided link to my other thread "GWM and Immich Integration (AI Image Search)".
Immich, the program I referenced is Free Open Source Software (FOSS). It is local, small, and lightweight; capable of being run on obsolete underpowered equipment and requiring minimal computational resources. It exists now and can be implemented now. That isn't the question.
The question is whether Chainer has the time, interest, and will to implement it on GWM.
The video linked below demonstrates the facial recognition abilities of Immich and details the specs of the obsolete pc the demo is performed on ($300 2018 Intel NUC Celeron).
"Immich stable vs. Zenni Optical - the facial recognition fight no one asked for" on Youtube
(3:26 of 4:46 total — I skipped ahead to 1:20)
I've been running immich for about a year and a half on our home network as part of our de-clouding efforts. It's great. We love it. But it is not the right tool for this site, nor does it claim to be or try to be. It's simply meant to be a google photos alternative, which it does swimmingly.
Its facial recognition is...ok. Fine for identifying you and your 30 closest friends and family. But it requires quite a lot of handholding and merging and whatnot. Very noisy. Finicky. Yes, it's resource-light; certain things are resource light for a reason. Not a big deal for a personal image collection, but maintaining it beyond a couple hundred people or so would not be fun. Things would get messy quite quickly. I would bet money that it would introduce more work for this site rather than less.
Another thing with the facial recognition is it can only match against what's known from the local collection. If you have photos of Bob and Alice, it can (with previous paragraph's caveats) detect them in newly added photos. But when you upload a photo of Carol for the first time, it's not going to know who she is. That's by design. The point of immich is that it not go out to the cloud. She'll show up in the People view as "Add A Name" until manually set.
I'm not aware of immich having any ability to detect AI slop or plans to add it, nor can I think of a use case for it having it, unless someone's firehosing photos in from somewhere, I suppose. But again, immich is meant to be a google photos alternative.
The party line on scaling is "we've heard of people storing millions of photos on their immich server," which as someone who's run high-traffic production sites before, would not have me jumping to inject that order of magnitude of photos into it. I know their goal is to officially support that volume, but I see no official word that they've met it yet. Photo collections will not scale up infinitely on a celeron. I don't believe our current photo collections would run well on a celeron. Immich has scaling info that talks about running parallel instances across multiple machines and all that, so it's quite clear that resource needs scale with collection size. This site's photo collection size and photo injection rate would need considerably more than a NUC.
And I haven't even gotten to the big one, which is that a handful of people accessing their personal photo collections on the home immich server several times throughout the day is a far cry removed from the traffic this site sees. I'm not aware of anyone running immich instances that serve more than the single or double digits, let alone whether the system can scale well to that, and I'm not aware of any goal to have the system handle that kind of incoming traffic.
I've been running immich for about a year and a half on our home network as part of our de-clouding efforts. It's great. We love it. But it is not the right tool for this site, nor does it claim to be or try to be. It's simply meant to be a google photos alternative, which it does swimmingly.
Its facial recognition is...ok. Fine for identifying you and your 30 closest friends and family. But it requires quite a lot of handholding and merging and whatnot. Very noisy. Finicky. Yes, it's resource-light; certain things are resource light for a reason. Not a big deal for a personal image collection, but maintaining it beyond a couple hundred people or so would not be fun. Things would get messy quite quickly. I would bet money that it would introduce more work for this site rather than less.
Another thing with the facial recognition is it can only match against what's known from the local collection. If you have photos of Bob and Alice, it can (with previous paragraph's caveats) detect them in newly added photos. But when you upload a photo of Carol for the first time, it's not going to know who she is. That's by design. The point of immich is that it not go out to the cloud. She'll show up in the People view as "Add A Name" until manually set.
I'm not aware of immich having any ability to detect AI slop or plans to add it, nor can I think of a use case for it having it, unless someone's firehosing photos in from somewhere, I suppose. But again, immich is meant to be a google photos alternative.
The party line on scaling is "we've heard of people storing millions of photos on their immich server," which as someone who's run high-traffic production sites before, would not have me jumping to inject that order of magnitude of photos into it. I know their goal is to officially support that volume, but I see no official word that they've met it yet. Photo collections will not scale up infinitely on a celeron. I don't believe our current photo collections would run well on a celeron. Immich has scaling info that talks about running parallel instances across multiple machines and all that, so it's quite clear that resource needs scale with collection size. This site's photo collection size and photo injection rate would need considerably more than a NUC.
And I haven't even gotten to the big one, which is that a handful of people accessing their personal photo collections on the home immich server several times throughout the day is a far cry removed from the traffic this site sees. I'm not aware of anyone running immich instances that serve more than the single or double digits, let alone whether the system can scale well to that, and I'm not aware of any goal to have the system handle that kind of incoming traffic.
Thank you for your input on this. I'm wondering if you're imagining a different implementation and use case than I intended. I believe you are. Did you read my other thread which goes into more detail? I could've elaborated better. Also, I listed my own concerns about Immich in the other thread.
Clarifications...
It doesn't need to be Immich, but it gets the conversation going for similar AI tool implementation even if the conversation ultimately moves away from Immich.
Yes, the use case for Immich as-is is wrong for GWM. But Immich is FOSS with an AGPL license.
From my other thread:
Immich has an AI image search tool with facial recognition that if integrated with GWM could streamline uploads and reduce manual management of duplicates, potentially automating the entire process. (Additional uses cited below.)
Immich is licensed under the GNU AGPL v3 License. It's free for Chainer to use, and he can modify it to be customized for GWM providing that he makes those modifications (obfuscated in ways specific to GWM internal structure and security) available to the public under the same license via GitHub or a similar repo.
I was never saying "Use Immich as-is site-wide, including image search.". I was saying "Immich is FOSS with an AGPL. Isolate the functions of the AI image search tool with facial recognition and modify those strictly for upload filtering. Mods only, not for user-access."
It can be applied for blacklisted models and for models that have requested to not appear on the site first. Next, it can be applied to the most uploaded models. If it was used for nothing more than that, it could significantly reduce workload on the Mods (which was the reason I suggested it).
I can't speak to the number of daily uploads to GWM, but I suspect there are peaks and valleys that use of a queue would level out. That would be something Chainer would need to write as part of the custom code forking necessary to repurpose only the AI image search tool with facial recognition parts from the Immich codebase.
Saying "This site's photo collection size and photo injection rate would need considerably more than a NUC." is a moot statement. I cited the Intel NUC example precisely as it was intended... as an example. I was not saying go out and buy a NUC. Obviously Chainer uses a web host that has much better equipment that GWM is already running on. Chainer can code a throttle for the resource usage if he chooses. The point was simply that for what it is, it's lightweight (even lighter stripped of superfluous features).
Since the purpose would be to screen the incoming uploads queue, this site's photo collection size is also moot. The objective wouldn't be to index every photo so it's user searchable. It would be to index an adequate sample across the collection of named models; and as explained above, not necessarily the entire (~98k name) collection, possibly only those requiring the most moderator screening.
You highlighted a point of confusion in my initial reply regarding AI slop and in how I placed the Immich link. The link overlapped to include my mention of AI slop. I've edited my reply to fix the link placement.
radek13's solution to address slop was to name more images. My reply used the same context. I didn't intend to imply that Immich screens for slop. I was saying here's a way to increase named images. Separately, I mentioned that (read: other) AI's exist which are specialized at identifying AI slop.
The top post right now is one that's up for debate if it's real or AI. There's no name given, so it's impossible to tell. I think posts without names should not even be allowed on the site. First of all, that enables you to post girls that have requested not to be featured here. Second, it lets you post AI generated slop, since without a name to go by, it's really hard to verify what's what.
I absolutely FUCKING HATE that we have to go trough this bullshit with AI taking over internet, but if we don't introduce some measures against it, this site can easily just get flooded by AI slop.