Stable Diffusion 2.0 Censorship

[darth2602]

Hello, Stable Diffusion 2.0 is coming, but I saw that there was NSFW censorship that was blocked "Greg Rutkowski" an other artist is this blocking only for the SD website version or is it also on the local computer and Google colab versions ?

Do you have any information on the subject?

[artificialguybr]

The 2.0 model dont have any filter.

[TheLastBen]

They limited the training dataset, so they don't need NSFW filter

[darth2602]

We are talking about stars, nude but especially the works of artists who allow for example to mix styles ?

Without models trained by the community to fix this, SD is almost useless ...

[darth2602]

I'm a bit disgusted by the way things are going, V2 is more advanced technically and much less artistically with the censorship and the artists/celebrities removed... I hope the community will quickly create models to fix this... because otherwise SD loses a lot of its interest...

the goal is not to make ridiculous parodies or obscene things but the nude is part of the art and to be able to make things "in the style of ...." or to mix the styles or to represent celebrities was the main interest of SD, under this new form, SD is almost useless...

We just created a Lamborghini with a dream body except we replaced its V12 engine by a fiat 500 engine because someone complained that it could go too fast

I'm exaggerating a bit but I think that without unofficial models trained by the community to fix it, stable diffusion is dead...

[nawnie]

i get being upset, but its not gonna take long for a dreambooth or some type of textual inversion to come out. as the project gets bigger its going to prune their data set to become more publicy acceptable, HOWEVER we can train at 768 now if we want even 1k IT WILL RUN but .. its gonna be like single subject training.

[TheLastBen]

you still can train 1024 pics on the 1024, here is an example of 1.5 trained on 1024, the result without highres.fix :

[darth2602]

About Dreambooth (with Google Colab) I would like to make my own model with celebrities

• what kind of photo pose is needed to define a person ? (face, right profile, left profile, close up, full view, ....) - what size ? 512x512, 768x768 , 1024x1024 I am on google colab (memory : high, processor : normal)

• how many pictures do I need to define a person ? - how many different people can I add in one training ? (there must be thousands in one official model...)

• if I understood well in the same training person 1 : albert, I must name all his images Albert 1.jpg, Albert 2.jpg, Albert 3.jpg and Albert will be the token to call it person 2 : Robert, I must name all his images Robert 1.jpg, Robert 2.jpg, Robert 3.jpg and Robert will be the token to call it etc etc

(I must choose particular names to differentiate them well so not a classic "Albert" but rather : Albert_DD, Robert_RD, etc etc)

• if I put 100 pictures, I have to multiply by 200 to have the number of steps ?

When I merge the models, what happens, does it add up or does it make a mixture? Can I train little by little and merge the models as I go along?

[TheLastBen]

use 768x768px images, training on people from the same gender will result in merging them, so keep the genders and the classes different.
you need 15-20 pictures per person, naming the pictures should be with unknown words like "hhyyuugg" instead of albert.

if you have 100 pictures, start with 10.000 steps (including 1000 steps (or 10%) text_encoder), then after that, keep training without the text encoder.

[gilroff]

use 768x768px images, training on people from the same gender will result in merging them, so keep the genders and the classes different. you need 15-20 pictures per person, naming the pictures should be with unknown words like "hhyyuugg" instead of albert.

if you have 100 pictures, start with 10.000 steps (including 1000 steps (or 10%) text_encoder), then after that, keep training without the text encoder.

Why is that when naming celebrities, we can't use their full name instead of nonexistent words?

[TheLastBen]

Some of them are badly represented and the existing knowledge may mess up the new training

[gilroff]

cool. i get you

[nawnie]

If yout to know about naming your model play with the tokenizer function in web ui and you can see how sd interprets your text, No hard limit on amount but every step causes changes to the base this means when I edit person a person b may take be affected, this is why not all celebs look right in sd, they probably did at one point. We can minimize this if our hard ware is good by increasing batch train size so we can do say 4 images before weighting to the model. But your biggest limitation here is that no 2 subject will take the same time to train, so with a handful you normally turn out good, but say over 10 and you may have to start adding weights to your prompt. Or compensait8ng if person b takes more steps to train chose less pictures or pictures that are closer in similarity, An easy cheat for naming is first 4 letters of first and last 4 of last . Numbers are disregarded in the sense of being used as file names to describe the subject, instead of John and John2 use Johntwo.. further information on file names can be found in the discussion referencing that, im on mobile but added a post to it last night should be near the top. There's alot under the hood here and sadly most guides lack in how to fine tune and only care about general use, once I finish my reference project so I cam have before and after pics i plan to male one.

…

On Fri, Nov 25, 2022, 4:58 AM darth2602 ***@***.***> wrote: About Dreambooth (with Google Colab) I would like to make my own model with celebrities - what kind of photo pose is needed to define a person ? (face, right profile, left profile, close up, full view, ....) - what size ? 512x512, 768x768 , 1024x1024 I am on google colab (memory : high, processor : normal) - how many pictures do I need to define a person ? - how many different people can I add in one training ? (there must be thousands in one official model...) - if I understood well in the same training person 1 : albert, I must name all his images Albert 1.jpg, Albert 2.jpg, Albert 3.jpg and Albert will be the token to call it person 2 : Robert, I must name all his images Robert 1.jpg, Robert 2.jpg, Robert 3.jpg and Robert will be the token to call it etc etc (I must choose particular names to differentiate them well so not a classic "Albert" but rather : Albert_DD, Robert_RD, etc etc) - if I put 100 pictures, I have to multiply by 200 to have the number of steps ? When I merge the models, what happens, does it add up or does it make a mixture? Can I train little by little and merge the models as I go along? — Reply to this email directly, view it on GitHub <#608 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZPYLSGBOS6JH6SIQYPDSATWKCLURANCNFSM6AAAAAASKUHYWU> . You are receiving this because you commented.Message ID: <TheLastBen/fast-stable-diffusion/repo-discussions/608/comments/4236412@ github.com>

[signalprime]

Does the community have a dataset with these removed CG artist images that we can all use for training? Would love to find that

[rektobot]

use 1.5 and just make a bunch of embeds or inversions of your favorites and use that. that is if those even will transfer over to sd2 🤷‍♂️

[nawnie]

Think we still crossing out fingers in a full realse before people take on the challenge. BUT.. I've started collecting already lol

…

On Mon, Nov 28, 2022, 1:55 PM signalprime ***@***.***> wrote: Does the community have a dataset with these removed CG artist images that we can all use for training? Would love to find that — Reply to this email directly, view it on GitHub <#608 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZPYLSCACZQJ6MHN5442HJDWKUE2TANCNFSM6AAAAAASKUHYWU> . You are receiving this because you commented.Message ID: <TheLastBen/fast-stable-diffusion/repo-discussions/608/comments/4256792@ github.com>

[signalprime]

@nawnie Good thinking, and I suppose we could always generate additional synthetic training data using the 1.5 model

[nawnie]

I'm just building data sets testing with 1.5 and will probably wait till 2.1 to bother with it to get an idea on what they are doing

…

On Fri, Dec 2, 2022, 12:50 PM signalprime ***@***.***> wrote: @nawnie <https://github.com/nawnie> Good thinking, and I suppose we could always generate additional synthetic training data using the 1.5 model — Reply to this email directly, view it on GitHub <#608 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZPYLSEOXTJRXB477BOSTLLWLJAI5ANCNFSM6AAAAAASKUHYWU> . You are receiving this because you were mentioned.Message ID: <TheLastBen/fast-stable-diffusion/repo-discussions/608/comments/4295080@ github.com>