tell us why. DPO is proposed as an alternative to RLHF which obviously was applied very successfull
tell us why. DPO is proposed as an alternative to RLHF which obviously was applied very successfully for LLMs
Ffine tuning configs for 6 gb, 8gb, 10 gb , 12 gb , 16gb and 24 gb
F
CI'll let you go read the paper and think
Dso you don't know the answer. alright.
Ci do, but i don't feel like hand-feeding you. and you trying to manipulate me into handing you the research you don't want to do isn't going to work
Dyou are making unsubstantiated claims that you know more about how Flux was created than the general public. anyone can read papers. you have to substantiate your claim.
Ci owe you nothing. i direct you, if you want a place to start your research other than arXiv, to all the posts on this discord ALONE where people talk about how many times they've tried to train flux for something and how it doesn't work. even Furkan himself has expressed issues with that. and that is all i'm giving you for free. go do the research
Cbe just as snarky as you like now, i'm done with you
Dblocked
Cblocking on discord is funny. you still see that the other person has posted something, but you can't see what. you can't tell if they're talking about you or not. blocking on discord is a good way to drive yourself crazy. and when you're the one that's tried to start an arguement, and then block the person you were trying to argue with because they wouldn't play your game, that says something about you to everyone watching
BWhats the best upscaler model that allows commercial use?
Ctopaz's models
Bcan they be intergrated into comfyUI?
Cno. they're propiratory and exclusive to topaz, and not free - but they are the best and you can use the output commercially
Byeah its not the output commercially im looking to do. Its for a the backend of a SaaS product.
Dcmon everyone, be nice to each other
D@Dr. Furkan Gözükara you can try flux finetuning with regularization and in the last epochs remove the regularization, may be the model get resemblance at the end but stays more flexible
Fit is an idea
Fmay work
Fi am testing each gpu config to get max speed
Fman this is taking time
Dflux-dev-de-distill is giving this error @kohya_tech is already informed and he said that he will fix it as soon a possible.
Fwow batch size 7 way faster than before
Fkohya did huge improvement
Ftraining takes under 2 hours haha
F3000 steps
F
Fah but my current speed test of batch size 7 not accurate
Fi need to give bigger
Fye made the images count 256 testing its speed again
Fye for 256 images not that fast 
Fso i guess best would be testing 14 images
D@crystalwizard You keep claiming that flux-dev can't be trained I was doing many tests with flux-dev-de-distill, I think that it can work. Its behavior on inference is completely different than regular flux-dev it works perfect without distilled cfg, what they did to model probably works, its behavior is similar to sdxl, if you prompt random you get random noise but if you prompt properly you get a perfect image, on regular flux-dev with distilled cfg you always get a coherent image, did you test it?
F14 images 200 epochs taking 3 hours
Fbatch size 7
Fso total 1$
Dgreat price!!
Cyeah. i spent a solid month testing flux
Fyep
Fnew configs
F
DTry this one https://huggingface.co/nyanko7/flux-dev-de-distill its behavior is completely different
Dthey did something to it, may be it surprises you
Ci know what they are doing, and it's not what they think they are doing.
Cplus - it's 100% impossible to put information back that's been removed if you don't know what it is - and they don't have the original weights.
D@Dr. Furkan Gözükara Kohya updated nyanko7/flux-dev-de-distill is training now.
DI agree you cant get that was deleted back, but it doesn't mean that you can't put new information on its place
Cbut that doesn't undo what was done, does it?
DI agree
Cif you have a bag of chocolate chips, i take them out, you have an empty bag. if i then put potato chips in it, do you have a bag of choclate chips again?
