Feel free to comment or ask questions here!
...
glad to see your back. fuck the civitai mods and reddit jannies.
Thanks! I am not going anywhere. I got into it for the love of the game and I am sticking around for the same reasons! I am posting on Tensor Hub as well.
Honestly, I think the issue is more to do with overreaching payment processors and legislators than any one hosting site. Still sucks, though. Will you eventually be posting all your loras here, or would that be too risky?
That is the root of the problem. Yes. CivitAI's reactions initially seem warranted until you look through the mods post. Yesterday, Theally had an image of Mark Wahlberg on his profile showcase. He posts images of himself and other mods. The rules don't apply to the mods over there... which will bite them in the ass. They also police video models like they are the end of the world but pay no mind to models like Flux.2 Klein which are just as capable of being problematic. It is an image edit model by nature. So you can take any F.2 Klein model trained on naked women, input an image of any woman and the output is that woman naked. But they don't police that model at all.
I will be uploading all of my models here. Existing and new models. It is going to take a little time though.
sup big dawg, long time first caller and all that. i have looked through every video lora that has ever been on civit (not just yours) and one question keeps popping up in my mind. some people clearly put in the effort to train and release loras that aren't garbage which takes some effort, and yet seemingly they don't put in any effort into the presentation as the example generations are liquid ass. i mean to say their prompts aren't creative at all which leads to uninspired examples and in a couple cases i tried those loras with better prompts and got better outputs. not always significantly better though , sometimes the model is clearly over trained and the one garbage example is all its capable of. conversely, your loras are great and your prompts are always creatively written so the examples look stellar. few people put in as much effort into their example gens.
so the question is, how much of what makes a good lora is the training (both source material and captions, as well as the algorithms and the selection of epochs that are trained just enough) compared to putting in a bit more effort to grind some gens with creative prompts?
p.s. fuck civit, playtime did nothin wrong. or maybe he did idunno i was out of the room
p.p.s your wan lora archive on hugface will be missed
It is all important. Crap in = crap out. The dataset has to be high quality and varied. The captioning has to be done correctly and written in the way the the base model wants to receive inputs )tags vs natural language, etc.). Plus, you need to understand what NOT to caption. Training parameters and epoch selection are VERY important. I always over-train and then test through the epochs to find what I feel is the perfect epoch... the one that does exactly what it is supposed to do and is the closest to being over-trained without actually being over-trained.
When it comes to prompting, of course it is always important. But the best trained models will give you a decent output with nothing more than the trigger word or phrase. For example, load up the titty drop lora and run a generation with "She lifts her shirt, exposing her breasts." and nothing more... It will give you a titty drop video. Everything else you prompt for will just give you more of what you want to see.
Also... the models will be coming back to HF. I am trying to figure out the best way to upload and organize them in the HF ecosystem. I have hundreds or loras, way more than were ever posted here before.
Have you thought about also keeping a torrent running with most of the models? Of all the options, torrents are the most resistant to censorship and it'll also mean several people have a full backup in case things go seriously wrong someday.
I have a full back up on a network attached server with dual redundancy... but I have thought about seeding a torrent as well.
Hey man, you are one of the best lora creators out there. No questions about it. But I would highly recommend not posting celebrity loras in general. They are one of the things that moves the ban hammer at lightning speed. This lora is particular might be even worse than a standard celebrity lora. While the actor is 18+, Wednesday Addams is depicted to be 15-16 in the show. It can be argued that this lora facilitates CP.
The community needs you and it would be a very sad day if you choose to leave because of regular bans due to smallest fraction of your loras.
Hey man, you are one of the best lora creators out there. No questions about it. But I would highly recommend not posting celebrity loras in general. They are one of the things that moves the ban hammer at lightning speed. This lora is particular might be even worse than a standard celebrity lora. While the actor is 18+, Wednesday Addams is depicted to be 15-16 in the show. It can be argued that this lora facilitates CP.
The community needs you and it would be a very sad day if you choose to leave because of regular bans due to smallest fraction of your loras.
I wouldn't say that simply having a Lora for a fictional character that is underage is illegal. I mean just look at the anime Lora's on Civit... Even base Illustrious 0.1 can produce illegal material if you wanted to. Until you do something dumb, I would say it's immoral at best to have some of them. It's what you do with the Loras that matter. However, yes, celeb content is not safe. Or any character that is simply too similar to the human portraying it, such as Indiana Jones.
Celeb content is best kept in torrents or private hosting. While Huggingface is best used for style, concept and fictional character Loras. (Such as Kermit and others) Better safe than sorry!
Hey man, you are one of the best lora creators out there. No questions about it. But I would highly recommend not posting celebrity loras in general. They are one of the things that moves the ban hammer at lightning speed. This lora is particular might be even worse than a standard celebrity lora. While the actor is 18+, Wednesday Addams is depicted to be 15-16 in the show. It can be argued that this lora facilitates CP.
The community needs you and it would be a very sad day if you choose to leave because of regular bans due to smallest fraction of your loras.
I wouldn't say that simply having a Lora for a fictional character that is underage is illegal. I mean just look at the anime Lora's on Civit... Even base Illustrious 0.1 can produce illegal material if you wanted to. Until you do something dumb, I would say it's immoral at best to have some of them. It's what you do with the Loras that matter. However, yes, celeb content is not safe. Or any character that is simply too similar to the human portraying it, such as Indiana Jones.
Celeb content is best kept in torrents or private hosting. While Huggingface is best used for style, concept and fictional character Loras. (Such as Kermit and others) Better safe than sorry!
I see your point. It also depends on the country, for example, in Canada images/videos that depict a minor in a sexual manner can be classified as a crime, regardless of whether a real child was involved in the creation.
Another point that could be relevant is that Huggingface also uses payment processors just like civitai and they have a zero-tolerance policy toward the sexualization of fictional minors, regardless of the art style.
That's what I mean though. Just because a Lora exists for an underage character, doesn't mean people have to use the model to sexualize. There are plenty of Lora's out there for a variety of characters , even on Civitai. Some are locked to sfw generations, some allow nsfw but you have to make sure the generated content is aged up or doesn't feature settings that have certain implications. School settings for example are almost always disallowed.
I think the main problem with Wednesday is simply that she looks too much like Jenny Ortega. There is no good way for regular folks like us to post content on Huggingface (videos, images) so you're not going to see this Lora used for bad stuff. So again, I think the main problem here is a real persons likeness.
Consider this another vote for torrents, you get it put up and I'll download and seed the whole thing long-term with ya. Even if you just do the ones most likely to trigger a response in torrents it would be a big step I think
Jenna Coleman incoming... testing now...
This one looks really good! I can't judge the Jenna since I am not familiar with her...
Why is this one 32 rank and Jenna 64? And what was the data set like for this one?
I am trying to figure out a minimum that will result in a good likeness. Right now I am getting good results with 20+ medium/closeup shots of the subject speaking. All are 1-5 sec clips that are then cropped to square (1024 or 512 depending on the source) and then resized at training time in Ai-Toolkit to 256x256. I have tried 256x and 512x and can't easily tell the difference. So I am sticking with 256 since it is 8s/it vs 15s/it on my 4070ti 12g.
I also supplement the videos with 12 HQ 768x1024 images in portrait to get fine face detail and body proportions. With those settings it starts to look/sound like them at 1k-2k epochs and fairly dialed before 3k.
This one looks really good! I can't judge the Jenna since I am not familiar with her...
Why is this one 32 rank and Jenna 64? And what was the data set like for this one?
I am trying to figure out a minimum that will result in a good likeness. Right now I am getting good results with 20+ medium/closeup shots of the subject speaking. All are 1-5 sec clips that are then cropped to square (1024 or 512 depending on the source) and then resized at training time in Ai-Toolkit to 256x256. I have tried 256x and 512x and can't easily tell the difference. So I am sticking with 256 since it is 8s/it vs 15s/it on my 4070ti 12g.
I also supplement the videos with 12 HQ 768x1024 images in portrait to get fine face detail and body proportions. With those settings it starts to look/sound like them at 1k-2k epochs and fairly dialed before 3k.
I am just trying things out... Seeing what works best. I can train at rank 64 and resize down if I don't see the benefit of the larger size.
Thanks for all your hard work @Playtime-AI . Looking forward to more fire loras from you!
Let me know if you have any ideas or requests...
Thanks for all your hard work @Playtime-AI . Looking forward to more fire loras from you!
Let me know if you have any ideas or requests...
Age slider or LTX2.3? I know i mentioned it before but i think the community would love it. the one you made here is amazing! https://civitaiarchive.com/models/2053301?modelVersionId=2323699