that is a good one to verify your training setup
that is a good one to verify your training setup
Dlemme see
KHow many epochs would you recommend for training 50 images for a 1024x1024 close-up face?
KSmall question: if I train my fine-tuned flux model, could I use it for inpainting, or do I need other parameters in koyas_ss to make it work a bit like flux fill?
KShould I do my training with flux fill instead of flux dev?
SYou can use your finetuned checkpoint to do impainting
Finetuning of flux fill isn't possible
Finetuning of flux fill isn't possible
SBetween 100 and 200. Around 150 is often the sweet spot with flux.dev, it depends on your dataset. I would recommend to save every 20 or 25 epochs, and compare on swarmui using the grid generation tool
Kok thx i will do this and follow the tutorial
K@skyrrr By any chance, are you familiar with face detection and Comfy UI? I'm struggling to create a workflow that would detect faces with Yolofacev8 or RetinaFace and then inpaint the face area, which would allow me to regenerate that part with my trained fluxfinetuned. But I can't find a good node. I tried installing “ComfyUI Yolov8” but I got an import error, so I guess it's not compatible. There's no information about this on GitHub.
�i'm scared to use this part of kohya, dont wanna waste a night of training and the model files werent saved anywhere...
�but would be practical for sure
Ki scared to haha
Ki didi'nt do the tutorial yet
Kbut i wish everything will work for my flux fine tuned
Kbecause time is money on cloudGPU :((
KIt's hell trying to find a simple node that allows me to simply only mask a face.
Ksomething like that
Kbut more simple i wish
Fand compare
Fit works right away as flux fill
Fyou can do inpainting
Fi tested once or twice never worked
KThx Furkan ! Do u kwn something good for me to face detection and mask ?
KI thought I had found the perfect thing for me, but apparently comfyui has decided to give me more errors, as always.
Kok i think it's maybe beacause python 3.10 only not above 
JHow many repeats and learning rate for 150 images at 20-25 epochs?
Fthat is why i still use python 3.10.11
Fmost widely supported version
Fuse our presets dont change
JI'm a tinkerer too 
JI guess I'll have to grep search then if I want the answer
SYou should follow @Dr. Furkan Gözükara method, it does wonders. Set repeats to 1. Train more epochs, at least 150
Four app can crop faces
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fand SAM can mask them
Fwe have SAM too
Kok thx i can't use this in comfyui ?
�what do you set guidance scale to when fine tuning a flux de destilled model?
筋I actually quite recommend an online ComfyUI running platform called Runinghub. You don't need to worry about ComfyUI environment errors. Basically, the latest technologies can be experienced or there are workflows available on it in a timely manner. The vast majority of them can be used directly and can also be downloaded. Although this may seem like an advertisement, I've actually subscribed to an annual membership myself. I use a 48G 4090 to run workflows and applications. If anyone needs it and it's appropriate, I can send the link.
筋
Fyou can use
�I just used default config. Will see how that goes. The 48gb one
�has anyone completed a wan 2.2 lora or fine tuning yet on a person? how do the results look?
CMostly problems with eyes...
I looking, thinking about right way how to train lora for wan2.2, specially what dataset to use. My doughter ask me to make for her LORA.
I dont know what aproach to do. Should I for HIGH noise use smaller resolution video clips, and for LOW NOISE use high quality photos.
Or maybe use same mixed dataset, videos and fotos, for both HIGH and LOW.
Hard for me to find any info wich way works best.
I looking, thinking about right way how to train lora for wan2.2, specially what dataset to use. My doughter ask me to make for her LORA.
I dont know what aproach to do. Should I for HIGH noise use smaller resolution video clips, and for LOW NOISE use high quality photos.
Or maybe use same mixed dataset, videos and fotos, for both HIGH and LOW.
Hard for me to find any info wich way works best.
�Epoch 80, flux krea fine tune. 25 pics
�Epoch 120 turned out to be best
�Used @Dr. Furkan Gözükara fine tune 48gb config
�Wut? My photo was 100 % sfw?
�Why remove it