Did anyone tried dreambooth Flux for a Style?
Did anyone tried dreambooth Flux for a Style?
DI solved it by just leaving 2 GPUs instead of 4.
D
Donly 2
Fcan be corrupted lora
Fnot yet
Fi am testing onetrainer
Fgreat
Fok first 8 training onetrainer completed time to test
�For all 99 Loras? 
Z@Dr. Furkan Gözükara once the dataset has been prepared, it doesn't matter if we leave these fields blank does it?
Zcause the folder structure for img is 1_ohwx_man
Fyou dont need to use there
Ffirst results with one trainer almost identical kohya
Fstill evaluating
Ztesting with gguf right now with trained lora in kohya, results are decent so far
Zi only got to train up to 75 steps yesterday cause my son wanted to play beamng, doing a full training round today
Z
Fnice
Zi'm so happy they came up with the idea of quantizing these models, now everyone can have a chance to do it
Flooks like scotastic rounding of onetrainer yields better than kohya
Zdo you have a tutorial on that? cause i have no idea what that is
Fwill do
Fi am asking kohya whether we have or not
Zand will it run on low vram systems
Fnope they dont have fp8 yet
Fso you need massed compute
Zgot it
Fthey have nf4
Fwaiting images to be generated
Zoooo
Fok nf4 works too
Fnot too bad
DI have this problem, in grid usage, the prompt is separated by the commas, and I need it not to do so, but I don't want to remove the commas.
Dadd || <- solution
CHello all. I recently followed @Dr. Furkan Gözükara FLUX Lora tutorial. I got a very good result when testing the Lora with Flux dev model, but not so good when testing with Flux dev NF4 model. Is it something I'm doing wrong, or is it just not compatible?
Zyeah it doesn't work well with nf4, surprisingly tho it works well with the gguf dev models @Cypriot79
Fuse this
Fnot compatiable
Ftry GGUF maybe better
Fi am preparing an expression dataset now
Fthis reddit people not beliving lol
Zdo the training samples work when training a kohya model, cause these images look nothing like me @Dr. Furkan Gözükara
Fnope dont trust samples
Fi dont use it
Zgotcha ok thanks
Zin swarmUI, what's the best way to inpaint your lora over an existing face?
Zor Rope might be better for that actually, now that i think about it
DI need a service or platform to be able to mount my API FLUX-dev + LORAS, and charge me per image generated. Since there are going to be few images per day.