Fstill unet only
Fclip l just arrived like 1 hour ago
Fi am waiting bmaltais
SOhh yeah ofc
Fi learnt the reason
Fflux has internal text encoding system that combines
Fgiven prompts + images
Fbut we are overfitting atm
Fi think fine tuning will fix it
SYeah I hope the clip_l training is even effective though
SThe lora extraction in kohya for flux is a bit broken sadly, results r poor and checkpoint itself is huge > 20GB
Cafter working on runpod and massed compute, i wanted to get my local machine up and running with flux. both swarm and comfy are updated. i ran the latest flux install from the patreon post and it installed (flux1-dev-bnb-nf4-v2.safetensors) into /swarmui/swarmui/models/stable-diffusion. the other flux models are in unet folder, and my comfyUI couldn't see (flux1-dev-bnb-nf4-v2.safetensors). i copied the model to the unet folder, then comfyUI could see the model but it broke. is there another way to install that particular model?
Ftrue it is getting better slowly
Fi think safetensors put into stable diffusion folder
Fbut let me check
SThe finetuned checkpoint has good results but hugee checkpoints so not really usable, but hope the lora extraction stuff gets fixed soon
Ccorrect. your installer put the (flux1-dev-bnb-nf4-v2.safetensors) into stable diffusion folder. but all the other flux models (flux1-dev.safetensors, flux1-dev-fp8.safetensors, flux1-schnell.safetensors, flux1-schnell_fp8_v2_unet.safetensors) into /swarmui/swarmui/models/unet -- and comfyUI couldn't find (flux1-dev-bnb-nf4-v2.safetensors)
Fgoes inot : SwarmUI/Models/Stable-Diffusion
Fuse latest downloader
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Fah i only tested with swarmui
F
Cah ok. yes i am using comfyUI workflows within the swarmUI interface
Fplease ask to dev he would reply
F
PHow will this ref t the training? And it’s available to use today?
HDoes anyone know if the windows running significantly slower than Linux bug fixed ?
Fnot on bmaltais yet
Fit is next research
Fwhat is your gpu
FA6000 speeds i posted on patreon almost should be equal to RTX 3090
Fso with rtx 4090 you need to get a lot faster
Cvia swarmUI discord: the comfyui back end does not support NF4 V2, only NF4 V1.
As far as I am aware, there is no developer interest in adding NF4 V2 support to comfyui, so it won't be available in SwarmUI.
As far as I am aware, there is no developer interest in adding NF4 V2 support to comfyui, so it won't be available in SwarmUI.
SFor someone with a 4090, what is the best model to use in a speed/quality tradeoff? I get confused with all the NF4/FP8/Q8_00 stuff
HI have a 3090
Jtry this lora now with a different prompt like photo of a ohwx woman, it should work, because “man” was not associated with your training photos. I am currently testing this stuff and it seems to be working.
Fuse FP8 directly
Fyou should get like 5.5 second it with slow
Fon windows
Fbut i get like 7.7 atm 
Fyep it is about flux internal text encoding
Fi am preparing video tutorial patreon post at the moment
Fso starting tutorial recording hopefully today
Hwhy you getting 7.7 it if its suposed to be 5.5
Hjuat wondering
Fwindows vs linux
Flinux is 5.5 windows is like 8 second it
Hdo you know whats causing windows to be slower
Hfor 3090