koha responded about the bleeding
koha responded about the bleeding
Fi will also upload so far best checkpoint training config
Flink pleae
DThe translation to English of the Japanese text is:
"Unfortunately, I don't have any ideas either. The lack of publicly available technical details about Flux is making the problem more difficult. I think we have no choice but to wait for the community's research."
"Unfortunately, I don't have any ideas either. The lack of publicly available technical details about Flux is making the problem more difficult. I think we have no choice but to wait for the community's research."
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
D
XViggle?
D@dsienra77 残念ですが、私もアイデアは持っていません。Fluxの技術的詳細が公開されていないことが、問題を難しくしていますね。
コミュニティの研究を待つしかないと思います。
コミュニティの研究を待つしかないと思います。
9/16/24, 12:39 AM
DDthicker beard lol
Fi wish flux pro was leaked 
Dlol
Di thought sd3 large would have been
Fi am pretty sure it is exactly same 23.8 gb
Fwe would directly use it with existing tools
D
hehe, may be that fixes the bleeding problem if is related to flux.dev being a distilled modelDwould be nice if black forest labs after releasing flux 2 pro it open source the version 1
Ftrying to find best fine tuning configs:
F
Fdmn fine tuning requires more than 16 gb :/
Fonly 24 gb gpus can train atm
Fhaha this is weird
F1 config uses 15.3 :d
Fnice 24 gb can train pretty decent speed
Fconfigs added
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Mhave 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