i tested fine tuning
i tested fine tuning
Ais the realism only on portraits or in general quality?
Acan you explain what you saw in brief
AIf I use the flux sigma model as checkpoint and load a Lora trained on flux dev . Will that move boost realism even more? what do you think @Furkan Gözükara SECourses
DLora trained on dev model will work worse, I guess
DI faced that with de distill.
NI rarely use Supir because I haven't figured out which configurations for certain styles are better, I did mass tests with models and I didn't really get a good result(what's better for realism, better for illustration, cartoon drawings). If there are such configs I would take a look at them
when I have some free time, I'll test it out
when I have some free time, I'll test it out
P@Furkan Gözükara SECourses can I generate iamge with swarmui on the already generated checkpoint while training is going on
A@Timson we need your wisdom
Tagree 
, I wish there were comprehensive guide like for flux, but I found supir guide lacking experiment data and considering how slow supir is (up to 2 minutes per image on L40S) it is hard to do research
, I wish there were comprehensive guide like for flux, but I found supir guide lacking experiment data and considering how slow supir is (up to 2 minutes per image on L40S) it is hard to do researchTsorry, coudn't find the question you are referring to
A@Timson
TI didn't explore Flux Sigma Vision, my impression was it is a model designed for special workflow with tiled generation, hence not compatible with most flux dev stuff
Tplus it's not a generalist model like flux dev, but model specifically trained for realistic close up portraits of men
T
T16 megapixel
JUsing massedcompute, how do I utilize two GPU's for fine tuning a model?
Jah wait found here https://youtu.be/-uhL2nW7Ddw?si=PeuMb2ddDNtDu8Se&t=1134
I assume the lora instructions are the same as for dream booth fine tuning?
only thing i worry about is the learning rate. if it's risky to calculate a new one by myself?
I assume the lora instructions are the same as for dream booth fine tuning?
only thing i worry about is the learning rate. if it's risky to calculate a new one by myself?
YouTubeSECourses
Unlock the power of FLUX LoRA training, even if you're short on GPUs or looking to boost speed and scale! This comprehensive guide takes you from novice to expert, showing you how to use Kohya GUI for creating top-notch FLUX LoRAs in the cloud. We'll cover everything: maximizing quality, optimizing speed, and finding the best deals. With our exc...
�Multi gpu dream booth training is only supported with 2 80gb gpu's
Jah shit okay
�@Furkan Gözükara SECourses Should I do a dream booth for this outfit - https://drive.google.com/drive/folders/1esrUCyCq9E1gu0WaKxoMFFzP1GvBi4Tj?usp=sharing ? I trained like 15 Loras on this with different settings and with all of them I have the same problem they are not learning the buttons correctly.
Google Drive
�Has anyone trained a dreambooth on clothes? If yes how were the results?
TI don't think flux is the right tool for replication of objects with 100% accuracy especially with tiny details like buttons and not expecting creativity and fluctuations at all
TIt may be possible with huge dataset (200+ images) with very good diversity and massive overtraining
Tyour dataset is 10 images with fairly poor diversity (same model, same lighting, same setting)
�I think that as well now after these countless trainings
A@Nitin Pareva you should train the pattern as a graphic instead and apply it to an outfit
Awhen I wanted to train a Stussy shirt i trained the stamp of the shirt
Aand it worked brilliantly
Aalso , I don't see a reason if flux can copy complex facial structures why it wouldn't do the same for a pattern
JAnyone knows how it works with dogs? If I fine tune a flux model on a person, where in some photos he's with his dog ... can i then prompt for "ohwx man walking his dog" and then it will be the same dog?
Tif it's less than 5% of training data you shouldn't worry.
Also you can edit|crop objects out if they interrsect your class or you worry about bleeding because of high percentage of traning data
Also you can edit|crop objects out if they interrsect your class or you worry about bleeding because of high percentage of traning data
DHow do you download the files you train on massed compute to your local pc? not seeing a way to do this in thinlinc
Aupload on hugging face
G@Furkan Gözükara SECourses thank you so much for the great work that you did,
quick question - to tine tune flux model using LoRA with GTX 3070 Ti 8GB which configuration and model size do you think works best in a reasonable time?
quick question - to tine tune flux model using LoRA with GTX 3070 Ti 8GB which configuration and model size do you think works best in a reasonable time?
Jhmm i mean I'd like to be able to generate a photo with this persons dog (not trying to avoid it). More just asking if the dog will be consistent in the final generated images if it's featured in some of the training data (say about 5% perhaps)?
JBest way is to upload to Huggingface using the python script and Jupiter notebook like Furkan explains in his tutorials
Fi couldnt test results in details
Fbut our config worked great
Fsadly i dont have too. we have 2 presets
F1 for loyalty
F1 for adding some more new details
Fyou can see in presets
Fyes if you have 2 gpus
Ffine tuning same but you need 80 gb gpus sadly
Fdreambooth better
Fand upscale generated images latent upscale
Fthere is a company we talking
Fif we deal i plan to research flux fill training
Fshould work fairly well
Fnot thinlinc