I ran Rank 1 Slow with T5 attention mask checked. I guess, the wait was worth it 'cuz the quality is
I ran Rank 1 Slow with T5 attention mask checked. I guess, the wait was worth it 'cuz the quality is amazing.
Fvery very small
Fit may
Fit improves slightly
Fit is accurate approach to do small test first
Fman i am gonna make a huge perspective dataset
Fflux is just amazing
F
Fwhen i add perspectives i expect my results to 10x get better
F
XWild!
DYou gave me an amazing style-training idea with this post 
DWhen does the style training vid come out, Doc?
DI can probably set up one next weekend now.
Fi dont know yet some people requesting onetrainer video :d
Fgot to sleep now hopefully see u later
DSweet dreams!
DSo, I just terminate the massed compute instance and it all goes away correct? 
DDude, you're gonna start a fire 
at massed compute
at massed computeXHow do i get a Unet folder path inside of swarmui so i can add flux model
DIts in models > unet
XIt’s not showing up for me
Awhats the prompt here? lol love it
PHello! whats the news on Flux Dreambooth training in Kohya on 24 vram ? possible already?
DCreate it.
Nare there any other samplers?
Nit's Flux
N
NAnd which upskellers are the best if you apply the Hires fix at the moment
Mfor generating is there a reason to use the fp8 model instead of the 23.8gb dev model? is it faster to use without loss of quality?
Mbtw this link in the kaggle notebook is not working https://huggingface.co/Kijai/flux-fp8/resolve/main/flux1-dev-fp8.safetensors so I'm downloading "flux1-dev-fp8-e4m3fn.safetensors" and " flux1-dev-fp8-e5m2.safetensors" separately from here https://huggingface.co/Kijai/flux-fp8/tree/main
NI tested it, 0 difference and moved to fp8
NErrors fail due to memory on the regular version of Dev
NI'm just training the lora on he
Moh you mean for lora training?
NYes, on instructions @Dr. Furkan Gözükara
DI have trained 3 custom Flux loras till now and everytime, I am getting scan lines in images. I use Rank 3 config (slow) with Kohya. Anyone else facing this issue? Not sure what I am doing wrong. I updated to latest Kohya version yesterday, but problem persists. Keeping Lora weight 0.8-1
DYou mean those vertical lines? Can it be an issue with something you're using to process the initial dataset?
DEven the bottom edge of the generation doesn't look clean, which makes me wonder.
DI have used 2 different datasets for custom Lora training till now. Each of 20 images
DBoth are producing such lines
DEven when I use different versions of same Lora saved at various epochs, getting scan lines
Not sure if this is because of overtraining. All images are 1024X1024
Not sure if this is because of overtraining. All images are 1024X1024
Dhere another example. Faint vertical lines are visible if you zoom in
M@Dr. Furkan Gözükara https://huggingface.co/OpenGVLab/InternVL2-8B/tree/main
this is the best captioner i found for flux, tested with 20 images, got my best results with them
this is the best captioner i found for flux, tested with 20 images, got my best results with them
M26B also nice, think could work on 4bit
DThhis is the top ranked in VLM models benchmarks
MYep even the 8b model tops many of the other ones
DDid you use any app or tool to load this model?
Mi manually used the huggingface space
Dnice, hope @Dr. Furkan Gözükara can create a set up like Cogvlm 2
DWhy does my terminated instance keep rising back up on Massed Compute? 