Error: '.ipynb_checkpoints' does not contain an underscore, skippin
Error: '.ipynb_checkpoints' does not contain an
underscore, skipping...
underscore, skipping...
Fnot error
Di put all in one folder and it worked without any changes, it stopped prior
Fyou are not using tolm right?
Dwhat is that/
Ftoml
Fif you dont know so you are not using
Dlol
Dmy loras turn out amazing with how i do it, but i read that having a primary file may help with faces
Di took the extra folder out and ran it, its currently working perfect now, but i want to figure out why adding a folder in "img" folder is doing that.
Dmy process i learned on reddit is to make two data sets of same images, 512x512 and 768x768, run each set seperate lora traing, then take the best of each and merge lora 100% and they are really nice. turn down weight to 0.4 and its amazing, you ever try that method?
Dby the way, i love our content, learned a ton from you. thanks
Fthank you so much
Fi never trained
Fbut you can achieve that in 1 training
Fput all images into same folder
Fenable bucketing
Fdifferent resolutions
SAnyone look into Stable Cascade yet?
Fyou mean training?
SJust in general, I heard something about them limiting it vs other models
PFloat16
PFloat8
Fi think no quality loss
Fbut float8 must be slower
P15.6s 30steps
Ffp16?
Pfloat8
Pfloat16 15.8s haha
Fwow nice
Fon llms
Fthere is huge difference
Fdid you see reduced vram?
Finteresting
PFloat 8 EulerDiscreter 30steps 15.6, Average image time: 15.6 Max. memory used: 7.1 GB with TyneVaeEncoder.
Fauto1111 changed output folder name :d
Ffrom outputs to output
Piam not testing in A1111
Psorry now i fixed fp8 in diffusers
P18.8s
P15.8s fp16
Fok this looks more accurate
Pwow 4.6GB
Pin cuda
Pwithout cpu offload
Pi will enable cpu offload to see how much will decrease
Fso we can apply this to instant id?=
Pno benefit with cpu offload for just one image, in fact it increased the time to 20s, memory remains the same.
PI think that because I can only apply float8 to unet and in cpu offload, unet is already the only one that remains in the cuda, so nothing changes if the memory consumption is lower than the total available of the video card.
Pwe need test,