i got to sleep. recently i am trying to sleep early since i have eye pain, i believe related
i got to sleep. recently i am trying to sleep early since i have eye pain, i believe related
ZYes have a good sleep doctor
Zgoodnight
Zwill update you tomorrow
Zappreciate the help!
Falso if you dont necessarily need sdxl
Fi strongly suggest flux
Fit is next level
ZI'm just trying my best to set it up myself, its really complicated and time consuming so I just wanna see if it works first
Zwill see about fine tuning later
Zjust following your stable diffusion guide
ZSent you a dm regarding some work, please respond when you can
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dyeh you only get one pair of eyes doc, you need to rest them
VNo way to do this through a replicate model is there, gonna need a custom pipeline?
D@Dr. Furkan Gözükara i got the block swapping to work by installing some more ram but the 12gb config is running at 40s/it. Any thoughts?
EHow much Ram do you have? Physical or virtual ram? It did not work for me
EDoes anyone know the best way to train the model to merge two persons? I have datasets of two different persons and want to create a unique mix. What would be the best way?
N"I'm training on 4090D with 500 images. The speed is 4.4s/it (using Rank_3_18950MB_9_05_Second_IT). When I'm training on A800 (80 GB VRAM), the speed is 3.3s/it (using Rank_1_29500MB_8_85_Second_IT). I think the speed difference is abnormal. What could be causing this?"
Fi think you can deploy any pipeline to replicate
Ftotally normal i got same speed
Fi had to purchase 2x32 gb today, my 4 stick 16 gb keep causing problems
Fi wasted 3 hours today
Fmin 48 gb phyiscal ram
Frank 3 is slow
Fyour speed totally good
Fi suggest fine tuning though way better results
OFor training styles, how important is it to have the training images in 1024x. It's a bit more difficult to get those in uniform resolution vs training a subject.
Fwell it is really hard to test
Fthus i suggest people to make 1 training with 1024x1024
Fand 1 training with bucketing
Oto compare? or as a step 2 on the same model/lora?
Fcompare 2 different trainings
Fand see which one works better
Fso you need to train twicew
OI know you've provided scripts/tools to id faces and crop etc, is there something similar that will just expand the canvas of an image to make it square?
Fit does that
Fif image resolution is big
Fand if can do without cropping subject
Fso that is why crop first
Fresize second
Oso if there's a series of portrait style images, it will expand the canvas to make it square, and I guess fill it with something?
Fif that image
Fhas person in distant
Fso that canvas can be expanded
Fif not it will crop
Fi think i added expand with white background too
NNShouldn't larger VRAM bring more speed? The A800 (using Rank 1) has more than 3 times the VRAM of the 4090D, but the speed is barely increased. Why is this? Is it because I haven't adjusted the batch_size?
Fno such thing. vram has 0 impact on speed if both fitting into vram
Fit is like case of a truck
Fbatch size can certainly make impact in your case