<@205854764540362752> what were your results with trying to train Flux on your reg data set? I'm att
@Furkan Gözükara SECourses what were your results with trying to train Flux on your reg data set? I'm attempting something similar
D@Furkan Gözükara SECourses please see here https://discord.com/channels/1102003518203756564/1149379539454140538/1294374243060027475 . I am not entirely sure this has been the reason your config didn't work, but it was the reason someone else's didn't work.
Fnever worked
LDthis is implemented in SimpleTuner now
Fchecking
Fhave you tested there?_
DNo, never used SimpleTuner so it would be a lot of effort for me to try
Dbut checked the code. they do exactly the same thing
Dother people are
Fi will wait kohya
DThe image is a high-resolution photograph featuring a woman with a stern expression, likely in her mid-30s, standing in a dimly lit, industrial setting. She has shoulder-length, straight blonde hair with a slight wave, and her skin has a light, natural complexion with a touch of blush on her cheeks. Her eyebrows are arched, and her lips are slightly pursed, conveying a look of determination or frustration. She is wearing a bright blue zip-up jacket, which is unzipped to reveal a hint of a purple top underneath. The jacket appears to be made of a smooth, synthetic material with a slightly shiny texture.
Civitai auto captioner - its free
Civitai auto captioner - its free
DLots of filler but otherwise reads solid enough
T@Furkan Gözükara SECourses have you tested the 6GB and 8GB configs for DreamBooth Flux training? For some reason i can't get it working on my 2080 super 8GB ram. I've tried both 6 and 8GB. Both crash as it starts the first step. Either out of memory or another error. I have no trouble on 3090. Also on the same PC with 2080 super, i can run fluxgym with same settings and it works.. But its not using DB method. "RuntimeError: CUDA error: out of memory
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with
steps: 0%| |"
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with
TORCH_USE_CUDA_DSA to enable device-side assertions.steps: 0%| |"
Fyou need 64 gb RAM
F- shared VRAM enabled
V
F712000 steps
Fyou really should watch tutorial
Fprobably made 40 repeat
V
Zestimated time remaining at the current iteration rate... 1245 hours, 53 min, and 37s
VIs that a typical time for the 48gb Dreambooth config at 100 images?
Fit is gonna take 1245 hours
Fplease read this article top to bottom
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
V@Furkan Gözükara SECourses I'm on that page atm. I ask because the hours are much longer then expected. To confirm the 48gb config should be 40 repeats and 200 epochs? (48GB_GPU_28200MB_6.4_second_it_Tier_1)
Fyou can see my timings
V@Furkan Gözükara SECourses I'm looking at the 48gb one. It says the repeat is set at 1 but the config is at 40. Should it be instead set at 1?
Frename your folder to 1
V@Furkan Gözükara SECourses I appreciate the confirmation. I would recommend double checking the epoch and repeat for your config, it might have mistakes in this area.
F
Fhow many images you have to train?
V89
Fok train up to 100 epoch at least
Fpreferably 150
Fand compare
VSo 150 epoch at 1 repeat or is higher recommended?
Falways 1 repeat with flux
V@Furkan Gözükara SECourses I see. That makes the results even more impressive! But the images are all of your face. Do you know if its recommended for styles like oil paintings?
Fit is totally same
Fi have style training
VThat's actually impressive. I always see in your videos pictures of yourself. 
Have you tried mixing it up with styles to see if you get more views/subscriptions?
Have you tried mixing it up with styles to see if you get more views/subscriptions?VI would recommend it. Training the images of the face is good for people who make celebrity loras etc, but most of us want to train on styles, concepts etc.
VYou can even do some on characters etc.
V@Furkan Gözükara SECourses Have you looked into flux rapid lora training? Both fal and civit.ai offer it.
https://education.civitai.com/quickstart-guide-to-flux-1/#rapid-flux-training
https://education.civitai.com/quickstart-guide-to-flux-1/#rapid-flux-training
Fi checked civitiai
