Did kohya work out Dr or did it still have bugs?
Did kohya work out Dr or did it still have bugs?
Fi found the bug
Fhopefully gonna start trainings today. but they keep fixing bugs adn adding new features 
Fi am also following every messages since there are so many parameters
Dcivitai is already exploding with flux LoRAs
ACan full size version be trained on 4090?
Fi think yes
Fi recently tested a lot of SDXL lora and most of them are trash :d
HLooking back now my SDXL Lora’s were very bad
FFyep mine too
AWhat made them bad Dr?
Flesser research and lora inferiority
Vwhy using sdxl and sd1.5..flux is here now..waiting for someone to make flux as optimized as sd 1.5 and sdxl so that i can have similar generation speeds as sd1.5 and sdxl....
Fgeneration speeds depends on model parameter size
Fbu for training hopefully i am gonna make 24 gb and 16 gb and 12 gb config
Vdoc what about GGUF Image Models ??
Vis it faster?
Vthere are so many GGUF models..which ones to download?
Vthere are so many models...flux schnell nf4...bitte nf4? ...salto nf4?? its progressing way too fast very hard to keep up..
Dwhat can you do on 16 gb?
Dabsolutely. for many even the preview pictures are trash. probably a result of civitai's "1 click trainer"
Dthe previews on the Flux loras however...if the previews aren't fake, they are really good
Dor better: what can't you do on 16 gb? low LoRA rank?
Nwhich flux version to download for 4090rtx?
�The highest quality one available Flux Dev I'd imagine
�along with the T5 FP16 text encoder for maximum power, the 4090 should handle it with ease
�I hope Nvidia hurries up and puts out a 48 VRAM consumer card for the 5000 series 
�or hopefully Dr. can figure out how the Chinese modded their cards and puts out a video tutorial 
Jif you want the best local then use Flux Dev. but not the fp8 version! plus the T5 FP16 text encoder as Alex said above. It takes a little while until everything is in memory and you can generate images, but after that it is actually quite fast, but if you change something at the prompt it takes a little longer again for the first time. I run it this way with WebUI Forge and SwarmUi. We just need more VRAM 
JReddit
Explore this post and more from the StableDiffusion community
JDoubt it's going to have 48gb vram, more like 32gb if they increase it.
Vi found the solution myself @Dr. Furkan Gözükara
VWhat's the fastest way to swap between loras for use in Flux dev generations? Like Pieter Levels's photoai or danny postmaa with headshotpro, how are they swapping between the loras for all their thousands of users on the fly? I've read in the past that Pieter at least uses replicate. Are they spinning up an instance, loading something like a comfyui environment, and downloading the lora to the server every single time? There's gotta be a faster/more efficient way
Fawesome
Fwell you have to load lora to model weights and then there is unload method on diffusers. i use it on instantid latest version
Fyou can see at source code
Fthey probably download loras and keep in ram for a while i dont think they have active thousands of users at a time
R@Dr. Furkan Gözükara will you be recording a tutorial for Flux kohya training?
Mhttps://www.letz.ai/blog/6_v2trainingguide
have you guys seen this, it train a image of a product to place the product in a image, i think it only can do this if the image used was generated by their model
have you guys seen this, it train a image of a product to place the product in a image, i think it only can do this if the image used was generated by their model
Mim trying to find out what do they use for this
M@Dr. Furkan Gözükara hey guys how can i train a model with product pictures in order to do inpainting on images with these products
i still dont understand how to do this, like placing a pop can in the hand of a model or something
i still dont understand how to do this, like placing a pop can in the hand of a model or something
M
Mor a brand t-shirt
�I haven't personally tested it before but I figure the logic would be something like using a dataset of a product with the backgrounds removed, so like on a plain background and also various angles of it, and no faces whatsoever in any of the images.
Mi see, any more details on it? ill reall ylike to have something like this
Mi mean how to mix the model and th eproduct together
Mis it lora so i can use 10 images of the product and assk it to generate something
SI didnt exprince it either , however, what about, give a try to, rare token word (xmycompanylogo) , and train 10 images just the logo, you want to see... after training to typo , "a woman wears a green shirt that xmycompanylogo is over it. " .. after train , you may want to extract the the training as a lora.
Fi am starting 8x A6000 GPU today for extensive testing
