You've probably heard of the new update to the Terms of Service that has added the following provision to the list of prohibited content.
Content that uses, reproduces, or is based on the likeness of real people - living or deceased - including public figures, celebrities, influencers, and private individuals, in any context. The staff still hasn't given me a clear answer on if a total ban on uploading photos of people for training was the intent of the prior added 18 U.S.C. §2257 disclaimer, though the official statement one can still use the trainer to train a LoRA of yourself or someone you have consent of indicates that wasn't intended. Similarly I haven't gotten an answer on how this new ToS clause applies to training non-"character" LoRAs, which arguably are not "[c]ontent that uses, reproduces, or is based on the likeness" of people with the on-site trainer that happen to use photographs in their data. (the lack of any answer makes me seriously suspect the staff is deliberately not wanting to put any statement on these subjects down so they can claim it is a particular thing latter as needed.)
The proposal
Since it would be far better than "can't use photos with people in any capacity ever" and would finally give me explicit confirmation in any direction, I propose Civitai adopt some set of rules/guidelines on the subject. I have come up with the following draft and encourage official adoption of rules/guidelines based on it.
When using photographs as training data, users shall follow the following to avoid training individual likeness
Faces of humans shall be obliterated from all photos used for training. Accepted methods include cropping to show only below the face, a solid color placed over the entire face, or strongly blurring the entire face.
No trigger words or description distinguishing a particular person from others in training data are permitted.
No one individual may be used for more than 20% of training data or 10 images total (whichever is less).
Non-photographic media 95 years or older (i.e., public domain worldwide) depicting specific individuals may be used as training data without obliteration (parts 2 and 3 still apply)
Photos of non-form fitting, fully concealing, outfits where no human skin is visible (e.g., astronauts with opaque visors, Darth Vader, "rubber suit" monsters) are considered to depict the outfit.
Notes on these:
1 could have some note about "Use of "identity censor" or "head out of frame" as applicable in labeling is recommended.", though that's getting beyond the actual subject and would need someone who knows more about "natural language" model making to give a counterpart for isolating it there.
3's exact numbers could be changed from those suggestions, but it's more than "one" to allow for clothing with multiple angles shown. Note for scale on numbers: 20 is generally considered the recommended minimum number for animated characters (below which model quality drops significantly. I'm told real people need even more data for a likeness to be taught) of pictures to train SD models (Flux is apparently lower) of most non-style type LoRAs types (thus only 4 images in a set of 20), and non-styles rarely warrant more than 60 images,
4 is a carveout for training styles of various historic artists. I think we can all agree including a painting of Lisa del Giocondo (who has been dead for nearly 500 years!) in a "Style of Leonardo da Vinci" LoRA is not a problem, let alone all the paintings where all information about their subject have been lost to time (such as Laughing Cavalier).
5 is a "nice to have" but droppable. This can't be said to depict the likeness of Haruo Nakajima, merely the Godzilla suit worn by Haruo Nakajima.
Why not use the suggestions feature?
Besides it being very, very slow, I would like public comment on the proposed. I will however submit it there as well.