I'm following the guide on setting it up but I already had anaconda installed and the miniconda exe gave me a ransomware warning on virustotal so I just changed the file paths in the .cmd to the already installed conda. However when I run it it fails to establish a connection to the repo. Any ideas? Making a firewall exception for conda didn't change anything.
Kevin Brooks
Link to the post explaining it. If the weight adjusting method is better then surely it can be adapted into the better syntax.
Cameron Baker
There is already one open and it hasn't hit the image limit what are you doong
Blake Powell
Previous thread
Cooper Hernandez
Any updates on some sort of training@home variant?
Colton Reyes
I don't think the syntax is better. You're free to search for the post, I won't do it for you. If you want, you can look at the code, it's available.
>It is very easy to conubt by glance how many ((((( you have. Hard disagree, especially if say we need to be between 1 and 10 for a complicated prompt. Also you have to be careful to match on the other side. Versus : 8, 1.7, 4 But I appreciate your work nonetheless user, ty
Xavier Ramirez
Post the link bro, I can't access pastebin.
Brandon Allen
I'm also interested on roughly speaking, how did you implement it better. Is there multiple ways to implement this feature (besides syntax thing)?
Alexander Jenkins
>promplets arguing about weights when they could be finding vectors
I strongly dislike the : syntax, and incidentally, since this has to work with the result of CLIP embedder rather than plain text, it will be substantially more difficult to implement. I probably could do it if I wanted, the dud who maintains the repo ou're all using, likely not.
Just in case someone is baited to respond, do not post the link, it's an autoban.
Evan Rogers
This place churns through 10 threads per day and I don't spend every waking moment of my time reading posts. I just weighed (heh) in the discussion about syntax. I'll check out the code.
Gabriel Taylor
From the code: # grabs all text up to the first occurrence of ':' as sub-prompt # takes the value following ':' as weight # if ':' has no value defined, defaults to 1.0 # repeats until no text remaining
Meaning it doesn't apply the weight to just what's immediately before the :n, but everything before it to the previous :, so it looks like the proper way to use it would be putting them on both sides of the term you want to adjust like >blah blah pink dress :big boobs:999 high heels, blah, blah, :blond hair:3 by greggy and the other guy
Jackson Jackson
I just checked his linkedin, what a fucking madman.
Connor Bailey
>Just in case someone is baited to respond, do not post the link, it's an autoban
You're no fun
Luis Anderson
the guy is a hero, give him a break
Brandon Turner
Ok so that works but it's simply unintuitive for normies.
Caleb Thompson
The :::: implementation takes each text separated by :, parses them, completely independently of each other, and blends results using user-provided weights (normalizing so that their sum is 1 unless you specify otherwise). My implementation parses one text just as original model, goes through results of the transformers layer, and multiplies values at positions corresponding to weighed terms according to weights; then I multiply the whole tensor by a value that restores original mean.
Christian Allen
oh yeah by all means prioritize a working thing over small details that could be hard to implement. I like your experiments. yup that's the way it works on MJ too, which mean copy-pasting prompts from the gallery would work too but I have a feeling nobody here cares about that.
>Meaning it doesn't apply the weight to just what's immediately before the :n, but everything before it to the previous :, so it looks like the proper way to use it would be putting them on both sides of the term you want to adjust like >>blah blah pink dress :big boobs:999 high heels, blah, blah, :blond hair:3 by greggy and the other guy And I know you are writing this without ever testing it yourself, because this leads to terrible results. Go ahead and try it at least once.
Connor Perry
see if you missed it
Christopher Davis
is it just me or are snakes kinda hard to generate properly?
Logan Rivera
ohhhh, interesting. It's really 2 different tools. The :: stuff can be great too in some cases tho (on MJ).
This does sound interesting. Would it be possible to implement both features (user selectable)? Or do you hate the fork too much? What made you two fall out anyway, I mean if it wasn't for him your fork would've stayed obscure. Hell I shilled it here, trying to get people to fork it. Now there's 40+ forks, with one popular one. Why don't you just bury the hatchet and work together?
Lincoln Sanchez
stop spamming sudo. you have wrecked your permissions. only use it after shit won't work without it
Juan Wilson
I have not once used sudo as this user. Hell, I even deleted the user, and then recreated it fresh, and still I get this error.
The #1 reason I can't accept :, is that : is a separator. It separates texts. And if the weight is one the other side of separatopr, it looks dumb and is unintuitive. "A :red 2.0: flag" would probably be a lot more acceptable than "A :red :2.0 flag", though I still like parens more.
It's not stealing if I don't get them. Hooliganism maybe?
I won't think we fell out, we never even talked. He made one pull request to me telling me to change saved pictures to jpeg, to which I said I won't do it by default for all users but this is already possible for those who want it via command line option, after that we didn't communicate.
Oh, and to answer the question, both can work at the same time. If mine in added to a codebase that supports :, I'm pretty sure no additional work is needed to make them compatible.
Isaac Williams
>He made one pull request to me telling me to change saved pictures to jpeg, to which I said I won't do it by default for all users but this is already possible for those who want it via command line option, after that we didn't communicate.
I mean, there's now a button (which is off by default) to save images as jpeg. Try and get in touch again once he wakes up, because working together will give everyone the best results, as opposed to the absolute trite that are the scattered reddit projects. This is the best UI out there, next to the commercial dreamscape one.
Parker Butler
The only vector you need is Greg Rutkowski, trending be his name.
