nphSi

/

Z-Image-Lora

like 116

Yes, the air has gotten very thin, and it's only a matter of time before HF pulls the plug here. Ultimately, being based in NY, they are subject to the NO FAKES Act.

However, HF does not have to be proactive in this regard (i.e., manually pre-screening all uploaded LoRAs to check whether they constitute an unauthorized celebrity AI). Instead, they only need to act if a) a celebrity (or their management/legal department) files a formal complaint with Hugging Face (Takedown Obligation), or b) they otherwise obtain knowledge of it (Risk of "Red Flag" Knowledge).

In that sense, the whole thing can still run via HF for a while longer, but as I said, it's only a matter of time...

In addition, there is of course the ever-growing risk for creators and distributors of LoRAs. While previously it was primarily the finished end product (the deepfake image or video) that was illegal, current legislation now directly targets the tools used to create them. Legally speaking, a LoRA of a real person is exactly that: a custom-made tool for cloning an identity.
But then again, we all know that pretty much every image AI out there—open or closed—already has tons of celebs baked right in. They can churn out pictures of them without needing any extra fine-tunes at all, even if it's usually mostly SFW. Technically, that means none of these models should even be distributed or offered anymore. Tricky.

Then again, nobody would actually use a LoRA for a deepfake nowadays. Face-swap fine-tunes for Qwen, Flux.Klein, and all that have gotten way too powerful, and they're so much easier to use if one wants cranking out NSFW deepfakes.
And you can't really ban those edit models, because that’d be like declaring a hammer a weapon and put it under such legislation, just because you COULD beat someone to death with it. But that's obviously not what a hammer is originally made for.

AFAIK if something is illegal by law the district attorney is forced to act on attention regardless of any complains by people maybe involved. Its criminal law and there is no more "Where there is no plaintiff, there is no judge." And a letter from a district attorney is a 100% killswitch...

AFAIK if something is illegal by law the district attorney is forced to act on attention regardless of any complains by people maybe involved. Its criminal law and there is no more "Where there is no plaintiff, there is no judge." And a letter from a district attorney is a 100% killswitch...

Absolutely. But even a district attorney would have to catch wind of this first and then reach out to HF. So it's just the hope the stuff on there flies under the radar for as long as possible.

Sadly that wont happen. The missing manpower to track will be filled by AI agents. Everything will be automated. Cloudflare said over 50% of all internet traffic is already done by AI agents.

@nphSi Thanks for your latest set of loras.

As for AI regulations and laws, I don't think the US has any federal regulations that oversee things like celeb fakes and it's more at the state level (https://www.cnn.com/2026/03/20/tech/white-house-ai-framework).

The US is taking a more lax approach to AI. Here's what the current national policy is on copyrighted material:

"Although the Administration believes that training of AI models on copyrighted
material does not violate copyright laws, it acknowledges arguments to the contrary
exist and therefore supports allowing the Courts to resolve this issue. Similarly,
Congress should not take any actions that would impact the judiciary’s resolution of
whether training on copyrighted material constitutes fair use."

(https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf?stream=top)

At least federally the laws aren't as strict.

Thats the reason why they push these AI laws. Copyright/DMCA does not hit hard enough. Some here in the EU call it "Digital Violence" and want to see it in the same league as Raping. And they get what they want until some years later higher courts fix the laws. Its the same game as always...

As for AI regulations and laws, I don't think the US has any federal regulations that oversee things like celeb fakes

2 days ago: https://www.blackburn.senate.gov/2026/6/blackburn-coons-bipartisan-bill-to-protect-individuals-and-creators-from-deepfakes-passes-senate-judiciary-committee
As I wrote, google NO FAKES Act.

As for AI regulations and laws, I don't think the US has any federal regulations that oversee things like celeb fakes

2 days ago: https://www.blackburn.senate.gov/2026/6/blackburn-coons-bipartisan-bill-to-protect-individuals-and-creators-from-deepfakes-passes-senate-judiciary-committee
As I wrote, google NO FAKES Act.

Still has to be signed into law, and who knows if that will happen with the current administration.

The last days i had some time to intensively think about current situation and i came to the decision to call this project finished and to slowly but surely stop working on it. It will stay online as long as HF allows it.
This will happen over the next few weeks, every now and then i will check suggestions/requests and may do small updates.

Reasons include

• OneTrainer is a slow dying horse, limited manpower for new models or even bug fixes.
• RunPod is a dead horse. Oversold and filled with slow or broken pods.
• Upgrading HW for newer models is impossible due to current prices.
• Z-Image is a dying horse, superseded by much better krea2.
• Upcoming AI regulation laws put me in an unpredictable risk.
• Personal interest in AI generally at a low.

Big thanks to all my supporters who helped covering training costs. I wish you all the best...

Really sorry to hear that - the Lora Index was my go-to spot almost every day over the last few weeks and months, but I can understand your reasoning for the most part. Personally, the lack of immediate pod availability bothers me not to much (as I described in the Krea thread). And in fact, my motivation has actually increased thanks to Krea, especially since the Image-to-Video quality has improved significantly too because of the better start frames. That said, Celebrity LoRAs are definitely hitting a deadline soon.

Anyway, I want to thank you again so much for all your hard work and effort. I’ve learned a ton here about training, and also realized that I need to look beyond my usual settings to get the high-res results I actually need. Thanks a lot for everything! Really appreciate it.

would you consider posting the datasets for others to train new models in the future? really appreciate your work

would you consider posting the datasets for others to train new models in the future? really appreciate your work

I am thinking about releasing all datasets as magnet torrent if HF pulls the plug here.

Thanks for all the LoRA files and I'll be sad when that page stops being updated but I can understand.
I am curious, when you train locally how long does it take at 100% processor load to turn a dataset into a LoRA? One day I really ought to learn how to do it.

An average 35 Images dataset (2500 Steps) is trained in about 70 minutes with OT. With a good RunPod 5090 its about 12 mins. (5 Loras/Hour/70cent).

Any chance you can write a tutorial for those who wants to learn how to do this ?

I prefer answering questions to specific topics. Understanding the process and what you do is important to have control over the results and to improve it.
But please open an own thread for that so others can find and use the info too.