SeaArt.ai Steals AI Content. Period.
For those who don’t know me, which I would assume is probably everyone, while in “real life” my name is Phil, in online communities I’m known under a number of handles. On social media, if I’m not using some variant of my actual name, I’m using kon16ov — a name I’ve been using since 1997. I use the kon16ov moniker on some creative sites as well, most notably DeviantArt.
For my AI-based endeavors, I can be found on Civitai.com as Soulbliterator. I’ve been a member there since October of 2023. This is the only site on which I publish checkpoint-merged models and LoRAs.
For reference, checkpoints are large piles of AI-readable stuff that direct the output when used for image generation. Merging two checkpoints with different strengths and weaknesses can, in theory, produce a merged model that will mitigate the weaknesses and combine the strengths. That said, a lot of testing goes into verifying that the checkpoint merges aren’t going to produce garbage images when each and every type of sampling method and/or schedule type another user will use to produce images on their own systems is used. LoRAs are smaller models that are fine-tuned to produce specific aspects of an image when generating, whether it’s background scenery, clothes, people, or particular styles of art. These also have to be tested.
There’s more to it than just testing, mind you. There’s a lot of data/image prep work that goes into refining a usable data set and yes, this is a gross oversimplification of the process since it differs based on what it’s being trained to do.
Testing involves, if done locally — as a lot of the content creators do — a pretty solid investment, technologically, and if you’re like me, you get no compensation to offset any costs. A local installation requires a CPU that won’t catch on fire when you push it hard, a GPU with enough power and VRAM to handle loading models, RAM for swapping — I didn’t work much on code, here, so what it does, even to me, falls under “magic.” That said, I have spent time tweaking code, mostly in Python, especially at the beginning of getting things up and running on my system. For reference, I have an AMD Ryzen 7 5800x-based system named Ryzengard with 64GB RAM, a 4060Ti w/16GB VRAM affectionally known as Plaidypus, and a pile of terabytes worth of storage with each drive named after a region in Middle Earth. Now, not all HD storage is dedicated to AI, but I have worked with LLM-based and Stable Diffusion-based installations, and the models range from a tight 2–3GB to a “monstrous” 6–8GB, each. I put monstrous in quotes because I’m working within my limitations. Some of the newer models that are released for the various formats can start at 24GB. The only folks that run the cards with sufficient VRAM for these models have plunked down some large outlays of cash — in the current market, probably upwards of $1500 for high-end cards with sufficient memory. These cards require a large power draw, as well, and properly powerful and efficient power supplies aren’t cheap, either. Basically, what I’m saying is that in order to play in the game, you’re fronting thousands of dollars for the computing power.
Again, I don’t know who all on the site is able to monetize this, but I know I don’t. I started looking into AI as a local installation a few years ago since, as a child of the 80s, I grew up with the tantalizing dream of AI and even wrote some short sci-fi stories revolving around it. That said, for me it’s a hobby. For others, they’re trying to make money providing tools others can use. So, as I’ve run down above, providing these tools takes a lot of time, processing power, and a little bit more time getting the tools up to the Civitai website and all that entails.
So, that brings us to the wild west-esque free-for-all that is the internet and the murky and ever-changing content rights enforcement landscape. There’s a site out there that is scraping the established and flourishing Civitai community’s content, and while not presenting it as its own, strictly, there are users set up with the same name and some of the same content without permission from the content creator. There are more than SeaArt, to be sure, but this is what I’m concentrating on since they’re, at least to the majority of the Civitai community, the most egregious “scraper” with very little recourse for the actual content creators to stop it.
I know, I know. I hear you. “It’s kind of rich for someone who is providing tools for others whose datasets are created or augmented by using images pulled from the web to be bent out of shape when it happens to them.” I understand the argument, but there’s a distinct difference. Most images used for training datasets are collected from public domain, publicly available images. While this may sound like justification, it probably is. This isn’t about how the models are created, since it can be argued about — and is — in much more detail elsewhere.
This is about the created models, posted by their legitimate content creators, being scraped and presented on a different website using the name of the content creator while not being the content creator. So far, there are two Soulbliterators on SeaArt, and neither of them is me. How do I know? First, I never set up an account. I would know. I’m me. It should be noted that I did create an account under the creative moniker “ Phil Ware.” Second, the users were created without using my email address, which is tied to the users I create. So, who is it? Good question. How did 2/3 of the models I’ve posted end up under these fraudulent accounts? Also, a very good question.
What can Civitai users do about this? Well, there are few options, and none are guaranteed results.
The ridiculous circular “legal-ese” on SeaArt’s site is headache-inducing, but, in theory, offers recourse. However, it looks like if you want to use their email-based reporting, you will have to basically list the original and the illegally posted version on a model-to-model basis. https://image.cdn2.seaart.ai/static/DMCA0305.html It’s a start, though not a really good one. The email to use: [email protected]
This is a titanic pain for any content creator on this site. I know that even for me and my meagre number of models, it’s a hefty number, and I can’t imagine how long and painful it would take for someone with a much larger number like the original poster of the article on Civitai, freckledvixon. At this point, with almost 1,800 models, that is a lot of effort to need to put in to prove that your created content is your own and was posted on another site without permission.
That’s the point, really. It’s not just that content we’ve posted on Civitai has been lifted and posted to another site under a fraudulent account. While that’s a large part of it, it is also that the vast majority of the models posted require a user to be logged in to Civitai to download, which means a bot or user created specifically for the purpose of siphoning models and their associated images from Civitai was created/used. To me, this speaks to intent, which is, honestly, why I am irked, along with a lot of Civitai content creators, as well.
That said, I was tempted to do it. I mean, how long could it take for, at the time, 120 models? For the record, a long time, which is infuriating inasmuch as it’s not my job, and yet it takes 4+ hours to put together the URLs for both the legitimate and fraudulent models in order to justify SeaArt taking either the fake user or the associated models down.
I’ve also been looking into the DMCA reporting on the DMCA website, but I’m not sure if this falls under the auspices of “copyrighted” or not. This is new enough that it’s nebulous, and it feels like nailing smoke to a tree. There are protections for content creators, though, and this, I would think, falls within those requirements to file with the DMCA. Here’s the DMCA FAQ about takedown requests: https://www.dmca.com/FAQ/Report-copyright-infringement
My current “solution” is that every model I train on Civitai, now, has a trigger embedded into it that says, “if this model is on any site other than civitai then it is there without permission.” It’s not perfect, by any stretch, but I have noticed that, so far at least, models with this tag have not shown up on SeaArt’s site. I don’t know if it’s coincidental with the added tags or with the increased spotlight being shown on SeaArt’s tactics. Maybe it’s a combination of the two. Either way, it’s something on which I’ll be keeping an eye.
Now, in the interest of “equal time,” something I learned a bit about as a DJ at a college radio station in the early ’90s, there are links to the Civitai-based posts/models listed on SeaArt’s pages for the same posts/models. While this is all well and good, it still skirts the major issue, and that is that the models from Civitai were posted without permission from the content creator. While a bit of an exaggeration, it’s a bit like museums of history posting a small placard discussing the country of origin from where the antiquities were… acquired.