3 ms·

Likely the latter. They hosted a lot of LoRAs and embeddings that help generate images of specific people. Users would post samples of those people nude. Hones

Likely the latter. They hosted a lot of LoRAs and embeddings that help generate images of specific people. Users would post samples of those people nude. Honestly, I'm surprised it lasted this long. Like that was clearly egregious. At least pretend to not be a deep fake porn generator...

Aeolun 1y ago
> At least pretend to not be a deep fake porn generator Why? Like, people are going to do it anyway. Whether in their minds or with some tool. Public or behind closed doors. That genie is very firmly out of the bottle, and it’s not going to go back in. At least when it’s on Civit it’s abundantly clear it is fake content. You can run any of these models on a consumer GPU.
- jazzyjackson 1y ago
  Because we live in a society and so get to have some influence on what behaviors we deem acceptable from our peers, and creating sexual images with the faces of your neighbors, acquaintances and even your foes is considered by some of us unacceptable, degenerate behavior Just because something is possible doesn't mean it has to be allowed to happen
- throwanem 1y ago
  I'll credit this argument when I see you publish a training set of selfies, or a model finetuned thereon. If you want to argue no one should expect to have anything to hide, you'd better do it from a glass house and in the altogether if you mean to be taken seriously, don't you think? It is a radical departure. The example is therefore necessarily yours to set, else you show yourself a hypocrite.
  - Aeolun 1y ago
    Just because I wouldn’t have an issue with random people generating porn of me, doesn’t mean I want to seek it out, see it, or facilitate it. It’s fundamentally no different from acknowledging that there may be a fanclub on the other side of the country with a stash of aeolun porn. As long as they’re not using it to advertise their services, and don’t tell me what kind of degenerate things they’re doing with it, what does that hurt me? Now that you’ve said it I’m actually mildly curious what would happen if I fine-tune a model purely on myself. But what would I do with it even if I made it? I can’t post it on Civit any more can I?
    - throwanem 1y ago
      I assume you could use it to produce generated images featuring yourself, which seems fine to me. But we make laws for the citizenry, not for you and me in particular. What news since diffusion models reached consumers suggests the citizenry can be trusted with tools this powerful? You can't buy a howitzer at Home Depot, either. These things are dangerous. I believe I recall hearing that people have died. Making the tools require some effort and skill to find seems warranted. Or did you think they cease to exist for no longer being so openly hosted? I'm sure such things are already traded in secret, and any traffic analyst with a sufficiently privileged perspective no doubt today knows much more about where and by whom.
pseudo0 1y ago
They always had a pretty strict policy against mixing any NSFW content with depictions of real people. Anything tagged as a real person would get run through a NSFW -detection API and would get automatically blocked if it failed. So no, no one was posting deepfaked nudes. Meanwhile I used a couple LoRAs to make JD Vance memes when that was the craze a few weeks ago, and now those are getting nuked... No fun allowed, thanks payment processors.
- godelski 1y ago
  Regardless, I still saw it happen quite a lot. Maybe I was seeing the images under different LoRAs. But I'm extremely confident I saw the celebrities in bikinis or other sexy settings. So let's be serious, any sexy setting, regardless of nudity, is making a certain implication... And sorry, but your NSFW detector is not going to catch that stuff (and it is going to frequently fail on even nudes). And did they apply the detector uniformly? Like you get the detector on Emma Watson, but does it apply to Hermine Granger? Because I definitely saw that situation reoccurring. Seemed to be extremely common with cartoon characters. Especially given the explicit nature of PonyDiffusion, mixing anime generation and realsm... Just turn of NSFW filters and go look at top models and top LoRAs. Nude content dominates the charts. Let's also be clear, there is a gender bias. I'd wager quite a lot of money that there are far fewer JD Vance nude deep fakes than many female celebrities.
  - pseudo0 1y ago
    They had the NSFW detection dialed up to the point where it would flag a bit of cleavage as inappropriate for real person depictions. They were very on top of this issue and really wanted to avoid any bad press over "deepfakes". Characters are a bit tricky. For example, Lara Croft could refer to the fictional video game character, or the character in the movie played by Angelina Jolie. They allowed the former, but not the latter. Or a model trained on the concept of "Hermione Granger" from fan art would be allowed, but not a LoRA trained on Emma Watson from the movies. That seemed like a pretty reasonable position to take.
    - 1y ago
rustcleaner 1y ago
Hopefully these LoRAs invigorate bittorrent again.

Aeolun 1y ago

> At least pretend to not be a deep fake porn generator Why? Like, people are going to do it anyway. Whether in their minds or with some tool. Public or behind closed doors. That genie is very firmly out of the bottle, and it’s not going to go back in. At least when it’s on Civit it’s abundantly clear it is fake content. You can run any of these models on a consumer GPU.

jazzyjackson 1y ago
Because we live in a society and so get to have some influence on what behaviors we deem acceptable from our peers, and creating sexual images with the faces of your neighbors, acquaintances and even your foes is considered by some of us unacceptable, degenerate behavior Just because something is possible doesn't mean it has to be allowed to happen
throwanem 1y ago
I'll credit this argument when I see you publish a training set of selfies, or a model finetuned thereon. If you want to argue no one should expect to have anything to hide, you'd better do it from a glass house and in the altogether if you mean to be taken seriously, don't you think? It is a radical departure. The example is therefore necessarily yours to set, else you show yourself a hypocrite.
- Aeolun 1y ago
  Just because I wouldn’t have an issue with random people generating porn of me, doesn’t mean I want to seek it out, see it, or facilitate it. It’s fundamentally no different from acknowledging that there may be a fanclub on the other side of the country with a stash of aeolun porn. As long as they’re not using it to advertise their services, and don’t tell me what kind of degenerate things they’re doing with it, what does that hurt me? Now that you’ve said it I’m actually mildly curious what would happen if I fine-tune a model purely on myself. But what would I do with it even if I made it? I can’t post it on Civit any more can I?
  - throwanem 1y ago
    I assume you could use it to produce generated images featuring yourself, which seems fine to me. But we make laws for the citizenry, not for you and me in particular. What news since diffusion models reached consumers suggests the citizenry can be trusted with tools this powerful? You can't buy a howitzer at Home Depot, either. These things are dangerous. I believe I recall hearing that people have died. Making the tools require some effort and skill to find seems warranted. Or did you think they cease to exist for no longer being so openly hosted? I'm sure such things are already traded in secret, and any traffic analyst with a sufficiently privileged perspective no doubt today knows much more about where and by whom.

pseudo0 1y ago

They always had a pretty strict policy against mixing any NSFW content with depictions of real people. Anything tagged as a real person would get run through a NSFW -detection API and would get automatically blocked if it failed. So no, no one was posting deepfaked nudes. Meanwhile I used a couple LoRAs to make JD Vance memes when that was the craze a few weeks ago, and now those are getting nuked... No fun allowed, thanks payment processors.

godelski 1y ago
Regardless, I still saw it happen quite a lot. Maybe I was seeing the images under different LoRAs. But I'm extremely confident I saw the celebrities in bikinis or other sexy settings. So let's be serious, any sexy setting, regardless of nudity, is making a certain implication... And sorry, but your NSFW detector is not going to catch that stuff (and it is going to frequently fail on even nudes). And did they apply the detector uniformly? Like you get the detector on Emma Watson, but does it apply to Hermine Granger? Because I definitely saw that situation reoccurring. Seemed to be extremely common with cartoon characters. Especially given the explicit nature of PonyDiffusion, mixing anime generation and realsm... Just turn of NSFW filters and go look at top models and top LoRAs. Nude content dominates the charts. Let's also be clear, there is a gender bias. I'd wager quite a lot of money that there are far fewer JD Vance nude deep fakes than many female celebrities.
- pseudo0 1y ago
  They had the NSFW detection dialed up to the point where it would flag a bit of cleavage as inappropriate for real person depictions. They were very on top of this issue and really wanted to avoid any bad press over "deepfakes". Characters are a bit tricky. For example, Lara Croft could refer to the fictional video game character, or the character in the movie played by Angelina Jolie. They allowed the former, but not the latter. Or a model trained on the concept of "Hermione Granger" from fan art would be allowed, but not a LoRA trained on Emma Watson from the movies. That seemed like a pretty reasonable position to take.
  - 1y ago

rustcleaner 1y ago

Hopefully these LoRAs invigorate bittorrent again.