i tested 10k images here with our reg images : https://huggingface.co/MonsterMMORPG/Big_FLUX_Fine_Tu
i tested 10k images here with our reg images : https://huggingface.co/MonsterMMORPG/Big_FLUX_Fine_Tuning_Experiments/tree/main
Faltough my attempt was general fine tuning
Fstyle may work
Adid it work ?
Fwell i had to use a very low LR
Fi did 5 epochs of 10k images
Fdidnt see significant improvement or anything
Fregular LR destroyed the model
Amaybe needs more
Awith very low lr
Alike 100 epoch
Fyep
Fpossibly
Awhat was the reason you stopped at 5 ?
Fi had to stop to do trainings for fine tuning tutorial 
Aah haha
Ais there much benefit to using H100 vs A6000 for this
Alike is the speed improvement worth it you think
Fwell
Aits almost 10x the cost
Fi think L40S and batch size 7 most sense making
Fbut on runpod speed was
Fway lower than I expect
Fmaybe it was that pod
Fi didnt test h100 speed
Fyou should test effective step speed
Aill test it later today and report back
VI do fine-tuning with dreambooth batch-7, I'm currently testing with 2e-5 and 1e-5 (4e-5 was indeed too high, but better too high than too low), with 4000 steps, I also activated buckets for multiple aspect ratio's (where each bucket must have at least 7 images). The first 10% of steps are warmup and the LR decreases with a cosine. It's looking like 2e-5 is working the best for my major concept.
Ahow many images total ?
VI'm currently doing stage 1 that has 556 images. The images in stage 1 has high concept and low details, most of these images are AI generated to really get the concept. For stage 2, I have 2551 images, these images have less of the concept but are high in detail and quality, which are mostly production grade photographs and marketing material for movies and such.
VI don't know what the config will be for stage 2.
Ahow are you ensuring each bucket has at least 7 ?
VI made a script. You want?
Ayeah sure if you dont mind sharing that would be great
VThe instructions are in the top of the file
fit_in_buckets.py12.22KB
Aamazing thank you bro, ill use it on my next test
Ahow did you come up with the idea to have min number of images per bucket ? did you notice it improves anything ?
VIt is because if you work with batches and you have buckets turned on. A batch will only contain images from a single bucket. So if you have buckets with 1 image. That means that it has to duplicate that single image 7 times to fill a batch, Making it basically ineffective.
Amakes sense
Awhat speed where you getting ?
AIve got H100 going now, its caching latents
Flet me check video part
Fi am still editing :d
F21.21 second
Fbatch size 7
Fby the way if dataset cant be divided into 7 exactly, you wont get very accurate speed
Aso H100, batch size 12 I getting 16s/it
A65GB GPU memory usage
Fwow nice
Falmost 1 second / it per image
AI can probably fit 14 batch size
