this is biggest problem of flux
this is biggest problem of flux
Fexpression test grids added
DI also think pose training does not work with Flux
DI've rented multiple GPUs since last two days and tried dreambooth training with various concepts. None have worked.
DI had surmised earlier that maybe flux_dev fails to learn multiple poses at the same time, so I tried training flux_dev_de_distill, but that did not work either.
DRight now, I am lora training a single pose, E.g., 'ohwx pose'. 15 images x 320 epochs (Rank1 settings & all images captioned as ohwx pose). If that does not work either, I'm not sure what will. 
DI guess Controlnets are still king when it comes to this sort of thing. 
But they don't seem to be as powerful with Flux as they were with SD.
But they don't seem to be as powerful with Flux as they were with SD. 
Fvery likely
F
DUndoutedly, the best character dreambooth workflow out there atm.
DDoc, had a question regarding dataset preparation! What if one artificially manipulated the background on the dataset to make it more varied (E.g., using photoshop or lightroom? Will the trainining pick up edge artifacts?
Fyep very likely
Fai picks those :d
DHaha, thought so. I've been very careful since that one time it picked up a mosquito lol
D
Dno doubt finetune might be better overall still, but you can definitely make a lora that has expressions without expressions in training data
Fbut wont be as good as fine tuning :d
DIf I want to add multiple people to a model (for instance me AND my wife) would I still use finetune or is lora better?
Dand how would I train it? Just add training images of both persons?
Ffine tune better
Fhave both in same picture
Flearns perfect
Fof course you need to caption in that case
Flike photo of ohwx man and bbuk woman
DDo both have to be in the same picture or can I have 10 pictures of ohwx man (captioned) and 10 pictures of bbuk woman (captioned)
D?
Dand also can I use not-normalized pictures in that case or is it bad?
Fi didnt test this so you need to test
Fbut people traines successfully
DWhat you are asking is currently experimental and includes some extra steps. In the end, you may or may not get the results you're looking for. If you are ready to fail a few times (like I have been failing for the past 3 days trying to do something experimental), I can tell you how lol.
Dyes please 
DAlright then, lets fail together in the DMs We come back once we are victorious lol.
We come back once we are victorious lol.D@Dr. Furkan Gözükara how does merging loras work? Any resources for learning that?
EI am trying to do a full Dreambooth finetune on massed compute with 2x A6000. I am using the config provided on patreon for 48gb gpu. I also set necessary parameters in accelerate tab but it keeps crashing due to OOM.
With one A6000 it worked just fine. Also is bucketing currently supported for Flux Dreambooth?
With one A6000 it worked just fine. Also is bucketing currently supported for Flux Dreambooth?
Fit is purely experimental
DI see
Fdreambooth requires 80 gb
Feach gpu
Fyou can do only lora
EI am pretty sure I did a full Finetune with 1x A6000 the other day
EAnd the model came out pretty good, full 23GB checkpoints
F1x uses 27.5 gb
F2x uses 77.5
Fbecause optimizations not working
Fi am talking with kohya he couldnt solve yet
EOh alright I see, so just 1 A6000 gpu is working for dreambooth currently
Fyep
Fit works even for 6 gb gpus :d
Fi made a very detailed publish just few minutes ago
