hi all, do we have any comfyui workflow to train flux on a character (like using dreambooth)
hi all, do we have any comfyui workflow to train flux on a character (like using dreambooth)
KNo. Not yet. Demands to much VRAM so noone has put any effort into that
KFlux trainings is done on rental-GPU's for now
Fokay i see, too bad!
FPS: you can run any comfyui workflow on rented GPUs on replicate: https://replicate.com/fofr/any-comfyui-workflow
Mhey doc, any toughts on how can i make the supir and vton app serverless after i deployed into runpod ?
Hinteresting.. can you run lora models as well?
FIs there a way to save your state from Massed Compute? It's kind of annoying that I have to reinstall all the custom nodes and models I wanna use everytime I boot an instance. Runpod's nice that I can keep a storage instance for cheap/hr
IHi guys, there's any tuto for loras on flux?
Sguys i tried to use topaz but a toaster runs better than my pc .. is there any colab or free alternative to topaz video enchancer???
Ihow long its the video?
S5 seconds
Ihmm then something its odd
Ais there a one click installer for ai toolkit?
Fi plan kohya tutorial soon for flux training
Fno idea you need some programming skills
Fbest is upload into hugging face
Fwhich i do
Fhopefully soon
Fthere is this app but only on windows i presume
FGitHub
Video, Image and GIF upscale/enlarge(Super-Resolution) and Video frame interpolation. Achieved with Waifu2x, Real-ESRGAN, Real-CUGAN, RTX Video Super Resolution VSR, SRMD, RealSR, Anime4K, RIFE, I...
KAnyone else having problems running the NF4 model in ComfyUI?
KI'm fully updated
KRuns in Forge though
KTrying to wrap my head around the differences between the weight types, models and VAE's.
I find that using the 327kb .sft file or using the 163kb .safetensor file produces results that are the same
Using weight type default produces way better images, but it takes longer. VAE-Decode takes tie with default. With the other ones it just takes a second
I find that using the 327kb .sft file or using the 163kb .safetensor file produces results that are the same
Using weight type default produces way better images, but it takes longer. VAE-Decode takes tie with default. With the other ones it just takes a second
KI find the Flux models to be highly unstable though. Sometimes it just produces blurry photos for no reason
KOnly thing changed between these three images is the seed that is incremented with one:
WTF is going on... ?
WTF is going on... ?
Fdo we have best practices on # steps per image and # of images to use for flux training?
KNo
TReddit
Explore this post and more from the StableDiffusion community
Fflux training i havent testerd yet
Fbut 30 steps yields best i presume
Fye flux will be super powerful for training
Fhopefully i will update my models downloader
WSo I downloaded the T5 installer after having success with flux.dev and flux.schnell and suddenly it seemed like it stopped loading the models entirely. My VRAM is not increasing, only system ram, and it's sitting on the log saying "User local requested 1 image with model 'Flux/flux1-schnell-fp8.safetensors'..." it's been 3 minutes..
Wsystem ram use is increasing now at 48gb.
Wok, it generated but load took 300 seconds. Any idea what's going on there?
WI have an RTX 4090 fyi.
Fit could be updating the backend
Fcan you restart and try again
Panyone tried training flux in ai toolkit? super easy to set up. But i dont real understand why XL captures likenes of a person but not changing the style, Yet SD3 and Flux trains both...not only the liknes but also changes style of images based on input quality
Fi havent tried yet
Fi am rather waiting kohya
Pi couldnt wait ))) was so curious. But yes a wait Kohya also. And your famousJson files with the best settings )
Fyep i will test all configs 
Fso many people doesnt properly test
Fso they never get best results
SDr Furkan i hope everything is ok, the control net for flux can work in forgue? and if it is, in what folder i must put it