Andreessen Horowitz Invests in Civitai, Which Profits from Nonconsensual AI Porn
404media.co
404media.co
> LoRA: Low-Rank Adaptation of Large Language Models
> An important paradigm of natural language processing consists of large-scale pre-training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full fine-tuning, which retrains all model parameters, becomes less feasible. Using GPT-3 175B as an example -- deploying independent instances of fine-tuned models, each with 175B parameters, is prohibitively expensive. We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks. Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. LoRA performs on-par or better than fine-tuning in model quality on RoBERTa, DeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters, a higher training throughput, and, unlike adapters, no additional inference latency.
An Emma Watson lora: https://civitai.com/models/123338/emma-watson-2010-sdxl?mode...
A lora to add motion trails: https://civitai.com/models/95680/motion-trail-after-image-or...
A lora to add a piano/black and white contrast style to the image: https://civitai.com/models/187675/piano-style-ebony-and-ivor...
If you view Civitai's list of all loras, sorted by most downloaded, you'll see a _lot_ of porn.
(Which parallels the role of nudes is other visual arts domains.)
Which, yay, big plus; porn slapping visitors on the face used to be a big deal.
Its still definitely there, and 404 clearly has a paid account since they posted about deliverately violating the nonconsensual porn policy for science with the paid onsite generation services.
That being said, the article is an awful sensationalist hit piece. It's as accurate as saying "Andreessen Horowitz Invests in YouTube, Which Profits from Schizophrenics Making Videos of Racist Tirades" or "AH Invests in Github, Which Profits from Software Banned in China and Iran"; true strictly speaking, but completely misrepresenting the purpose of the service.
Being an investor for some company does not make one responsible for the company breaking the law. Even less so for allegedly breaking the law.
How was this not buried instantly?
They dug and dug, and found something to attack them with. A connection by investing into a company allegedly doing something illegal and deeply immoral.
Let me put extra emphasis: Horowitz isn't allegedly doing something illegal and immoral. A company he is investing on is allegedly doing something illegal and immoral.
We're supporting this lynching campaign by tolerating it and giving it attention.
What, if anything, is wrong with this?
https://en.wikipedia.org/wiki/Andreessen_Horowitz
(And that's stated in the very first sentence of the article being discussed here.)
I quickly closed the page when I realized the article needs some sort of membership to read, that's likely how I missed this.
I'm less sympathetic with companies, but what I said still applies somewhat. This is a Civitai problem, not a Civitai investor problem.
I might argue this could be good in that this sort of stuff will become so commonplace that it won't be a big deal and people will learn (be forced to shrug it off). Up until now-ish, the scarcity of celebrity nude leaks and revenge porn type stuff made it somewhat of a forbidden fruit. I like the idea of destigmatizing this sort of stuff so that it loses power.
That said, I'm sure there will be some casualties along the way. Eg vulnerable teens who are picked on by mean girls / boys, semi-celebrities / public figures who are targeted by online trolls, etc. I hope that the prevalence of this sort of stuff enlarges the conversation around this, increasing the reach of strategies to deal with this stuff psychologically, and significantly reduces it's impact.
I don't know many celebrities or politicians, but I suspect that after they see a few weird / porn related photos of themselves, they stop caring.
Thoughts?
But I am concerned, for a different reason.
Civitai's enshittification potential is sky high. They are hosting loads of enormous models and images for free, on a very spiffy, heavy website. There's no way thats sustainable.
I share these same concerns.
What this space desperately needs is to have models shared via torrent. As it is today Civit.ai has had problems with creators decided to take down their own models and suddenly people who were using them have no place to find them.
I suspect we'll see more and more lockdowns on what models are deemed "appropriate" and it would be nice to have a decentralized alternative.
Not to mention torrenting is the only rational way to sustainably host such large files at no cost.
And there already are torrents of SD models. Just filter by the .safetensors filetype.
If would be nice if their hosting was decentralized though. Users seem to hoard LoRAs and merges in personal collection anyway, mind as well help distroute them so Civit doesn't have to enshittify so quick.
"Oh no, the model can generate porn!" "Oh no, it can generate realistic photos of people!" "Oh no, it can generate violent or racist images!"
I'm so sick of people arguing that we should just shut down AI research or commercialization because it could be used maliciously. Think of how many grocery store products could be used maliciously if someone wanted to. Should we not allow people to buy bleach? Should we not allow knives to be sold?
People need to get over their misgivings from this new technology and start taking accountability for that fact that models can be abused like any other tool. That doesn't mean we should just not give anyone access to them.
Ya know, take one for the team so to speak.
No, its misleading by selective focus (like the previous article it references) to get people to come to that conclusion themselves, a much more effective persusasive technique than the explicit demand because with the explicit demand people assume your evidence is selective to support the demand which may have ulterior motive, while they are less likely to do so if your presentation is superficially merely informative but actually selectively crafted to motivate a particular conclusion.
The people who download a checkpoint and a LoRA (a special kind of sub-checkpoint or "model modifier" if you will) of some celebrity are the ones who would be generating something like porn using that celebrity's likeness. Not civitai.
In this instance civitai is providing tools that can be used to generate unauthorized porn just like Home Depot provides tools that can be used to make illegal weapons.
Is it illegal to make a LoRA of someone's likeness without their permission? No. Because until it is used to generate images which are then published it's like a gun in a safe, doing nothing... Just sitting there.
There's no federal laws protecting likeness anyway (only certain states). So civitai has no real obligation there.
Is it ethical to make a model of someone's likeness without their permission though? I don't know. To me, it's akin to taking a picture of someone in public... If they're in a public space they're "broadcasting" their likeness.
You can make a really great LoRA of anyone's likeness with just a handful of images. If you're skilled you can probably do it with just one image (using some controlnet stuff).
I mean, do people have a right to privacy? Is there some line between having privacy and infringing on someone’s freedom of expression through AI content generators that you would accept?
And if safeguards cannot be put into place by design into your tool, then that is malicious design
Right. So can paintbrushes. So can guys with deer antlers scratching on cave walls.
Controlling likenesses in AI was the whole point of the SAG strike.
What if they used Photoshop instead of drawing from a reference?
The fact that a small problem exists does not negate the existence of a larger problem.
Yeah, widely publishing or making money off of someone's likeness is a whole different can of worms. I was thinking for strictly personal use.
Right now CivitAI is basically Stable Diffusion social media. It doesn't seem profitable at all, even of they are selling user data or getting subs or whatever.
But if they start charging for downloads or whatever, they may cross a line into making money off likeness, and not just "hosting user content" like Facebook and Twitter and any oldschool image hosting service gets to say they do.
Yes, commercial use of likeness is also an issue, and it may or may not be violated simply by distribution of something use the likeness on a commercial website like civitai.
Revenge porn laws are generally criminal, and as such unaffected by Section 230.
I think they had 5 or 6 demands.
No, it wasn't.
It was one important issue in the strike, but there were others (streaming residuals were a big issue, for instance.)
Who gets to decide what should or should not be necessary? Do you think your opinion about this is the majority view by people?
https://en.wikipedia.org/wiki/Personality_rights#United_Stat...
So, the USA-specific answer is depends on the specific US state(s) whose law relevant to the action in controversy.
There are countries with national rights in this area, but the USA is (and your source highlights this) not one of them.
> What if I look identical to Taylor Swift? Do I lose my right to free expression?
Taylor Swift's corporate legal team would already do a pretty fine job of excising that right from you. No additional legislation needed.
Yes.
> What if I look identical to Taylor Swift?
Yes, still.
Now, if you are trading commercially on the appearance similarity in a way which presents, either explicitly or implicitly on your images as Swift’s, then you open yourself up to right of publicity claims in some jurisdictions, and the same may be true of revenge porn laws in some jurisdiction, even without the context being commercial.
I'm not entitled to categorically own/forbid using a look. That's nonsense and leads to self-inflicted quandaries: How do I know a video of unknown provenance contains me, not a dead ringer that gave consent? How different must a depiction be to not require my consent? 9 pixels? 30%, whatever that means? At least an eye color change?
It's impossible to consistently enforce, presumptive, and effectively thought-policing a concept. In short: it's absurd.
> It's impossible to consistently enforce, presumptive, and effectively thought-policing a concept. In short: it's absurd.
I mean, come on. It’s fine to disapprove of the law, but this isn’t some uniquely difficult thing that the legal system couldn’t possibly handle. It’s certainly nowhere near the level of complexity and ambiguity of, for instance, criminal fraud law, where things like the intent of the accused and the “reasonable person” are routinely crucial elements.
But it is neither novel nor unique to AI.
What do?
I dunno about early access.
Huggingface also has porn-focussed trained stable-diffusion models. [0] They also have celebrity-trained models. [1]
[0] e.g., https://huggingface.co/WolfAether21/NSFW-STABLE-DIFFUSION-LO...
[1] e.g., (yes, poor quality from the samples) https://huggingface.co/lora-library/12jenniferl
The highlighted bounty feature is technically accurately, but misleadingly, described. Yes, people can and do post requests for models of real celebs there, but they also post requests for fictional character models, for concept models not tied to particular real or fictional characters, for technical assistance with model training, for people to generate showcase images for existing models.
[edit] The fact is we've had this kind of content for a long time, and we've developed techniques for identifying and removing it. There's great nudity detecting ML models. I suspect the choice was made in part ideologically, and in part to drive sensationalist coverage, and it looks like it's working.
No, they screen, remove, and ban people for policy violations. (And that has improved over time, at least, the violations that have slipped through have dropped over time.)
What isn't a policy violation is separate things which could be combined together or with other things to make something that would be a policy violations.
The dot in “CivitAI" is between “AI” and “com” if you are referring to the domain name, or just “there isn't one” if you are referring to the name.
Civitai is just a popular place to host stable diffusion models / LORAs / textual inversions, not some kind of fake nudes marketplace. It's the huggingface of stable diffusion.