have you tried the extracted lora
have you tried the extracted lora
Myou need to up the weight on them to be usable
Fi didnt
Fbut fine tuning definitely better
Fstarted 8 more test need to sleep now 6:44 am :/
Fthe lowest vram is 15450 mb
Fsupposed to be no quality loss
Pfinaly! finetune time! cant wait )) LoRA is defenetely not good enougth...
Dspeeds are very resolution dependent. I am getting 1.1 it/s with 512 but around 5 it/s with 1024, training Flux Lora. Would you expect the parameters like LR to be the same for both, or have you seen that you need other parameters at higher resolutions?
DAsking because if it's the same, you can make all your adjustments and tests using 512, and only go to 1024 for the final run
LHey guys i have question, i. trained lora with my dataset, and it went great, i saw there is some lora's out there like photorealistic loras that people did, can i train again my dataset image with those lora? or you cant do that, if i understand correctly lora it's to train specific things in the model right?
Dyou could, by merging the lora into the Flux checkpoint first, and then train on that. Whether this improves the result for the Flux realistic Lora is open to debate, but your theory is correct.
I have trained in the past on an SDXL checkpoint merged like that, because I knew I wanted to use it with another Lora that negatively affected my lora, unless trained like that
I have trained in the past on an SDXL checkpoint merged like that, because I knew I wanted to use it with another Lora that negatively affected my lora, unless trained like that
DI don't know if there are already good merging tools for Flux (let me know if you find one), but there are for SDXL
Loh ok i understand
Lthank you
Dnp. that's actually something I plan to try, with the Flux realistic Lora
Dwhat speeds do you get on 16 GB vram? does it require the very slow model splitting?
DI have tried the new configurations for lora training and the results are getting better, thanks!
Ji still see no reason to train a full fintetuning and get a 23GB model compared to a few MB for a LoRA.
I already have around 20 LoRA's' and the results are very good. But will it be possible to extract a loRA's
from the finteuning model? because that could be interesting.
I already have around 20 LoRA's' and the results are very good. But will it be possible to extract a loRA's
from the finteuning model? because that could be interesting.
Dwhether finetunes are still better than direct Lora/Dora training with Flux I think is currently unknown. With SDXL it was the case. It could well be that a high-rank Dora is now good enough and you won't see a difference anymore.
Fsame parameters
Fyou can merge that lora into base model
Fand train your lora on merged model
F16 gb is uber slow
F39 second / it
Fwe need fp8 training
EGitHub
Official code for PuLID: Pure and Lightning ID Customization via Contrastive Alignment - ToTheBeginning/PuLID
FEOh sorry, got it.
Dnow Flux training became fun instead of a chore. you can run 3000 steps in 1 hour, which is equivalent to 6000 steps on SDXL because SDXL benefited from regularization images a lot. so Flux is now faster to test than SDXL, because SDXL couldn't handle low resolution training 
Swhat means this Dr Furkan . ican sue 512 or 768 image for trianing in flux and not loos quality?
Floses a bit
Fit is a trade off
DIs there anyone who trained a LoRa on Schnell and succeed? I trained a lora on schnell with my perfonal photos but it didn't learn anything.
Fi trained
Fit is bad
FReddit
Explore this post and more from the SECourses community
DI was trying to find a fast generation method like this and thought it might be Schnell, but apparently Schnell is not good for personal photo generation: https://x.com/levelsio/status/1833562269281616287
Fthis guy sells 100x cost of course he can use very expensive gpus
DI am currently using Flux dev and it is not that fast. Do you have an idea Dr.?
Fhe said his profit margin is 99% 
DI mean if what he uses is just a Flux Dev but fast gpus, thats ok
Frtx 4090 is already 5 second
Fon swarmui
Fwith --fast
DBut I thought he might be using another pipeline
Dcan you give a little bit more details on the generation settings, like resolution, refiner, upscale etc?
Fjsut 1024x1024 20 steps