<@205854764540362752> Will you make a video to tell us how to train HiDream by using Dreambooth in M
@Furkan Gözükara SECourses Will you make a video to tell us how to train HiDream by using Dreambooth in MassedCompute Cloud?
Fyes i plan to. the things that comes never ending :d
TThanks very much, I'd like to use HiDream but without Dreambooth, not much use for me for now.
I know you are gonna make it ! Thanks @Furkan Gözükara SECourses
I know you are gonna make it ! Thanks @Furkan Gözükara SECourses
SDo you think it worth it compared to flux dev finetune ?
TFlux has too many internal censorships
�Has anyone fine tuned on pixelwave checkpoint?
�Or made lora
Fi did
FMedium
Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…
Fwhat i care is multi concept
Fif works or not
Fotherwise for single concept flux pretty much amazingly working
EHello, I want to train sdxl koyha_ss on massed compute with A100 graphic card, I use the Tier1_48_GB_Faster_v2 config but I don't know what epoch I should set. I only have 8 images and I use all 5000 reg images. From what I understand, I should aim for 3000-5000 total steps. So I'm currently trying with the default settings of the configuration but setting 400 repeats so I have about 3200 steps...
- Is 400 repeats with 1 epoch and 8 pictures good?
- I set a batch size of 4 since the A100 gpu has 80gb of memory. Is this right?
- Could I optimize the config better to be faster or more efficient for A100 gpu?
EAlso, on massed compute, I see
Ah and if it's only 8 portrait images, do I want captions? From what I understand no
Thanks
Could not find TensorRTAh and if it's only 8 portrait images, do I want captions? From what I understand no
Thanks
Fnot important
F8 images really low
Fso go up to 300 epochs
Fsince it is sdxl
Fmake sure to use reg images
Ffor kohya
Fyou need to test repeating to be able to use class images
F400 repeat 1 epoch sounding good
Fmake sure to save once every n steps
EOk, and is the tier1 48gb config optimal for A100 graphic card with 80gb or I should change something to make it better and use more available memory? Batch size around 4-7? Do you recommend onetrainer finetune over kohya_ss dreambooth?
Thank you for the quick answer
Thank you for the quick answer
Fi dont recommend change anything it is best quality 
EOk
I'll try to give results here if I find the best sweet spot for my number of training images!
I do think it could be useful to have somer kind of pinned message with recommended setups of epoch #, repeat # and such specific to the # of images, or at least a ballpark close to it
Like 8 images: 400 repeats, 1 epoch
I'll try to give results here if I find the best sweet spot for my number of training images!
I do think it could be useful to have somer kind of pinned message with recommended setups of epoch #, repeat # and such specific to the # of images, or at least a ballpark close to it
Like 8 images: 400 repeats, 1 epoch
Fsadly there isnt much but i usually recommend 200 epoch
Fhowever your train image count too low
Fmore is better
EYeah I only have 8 images, I'm hoping I can finetune a model and generate decent enough images to be able to train with more images... Not sure how that will go lol
�hi, for quick trainings I often use civitai. I remember you talked about rank of lora in kohya, here it is both rank and alpha, and not sure about these. The default seems to be the same for both flux and sdxl trainings: Default is also not adafactor, but the Adam...any recommendations here? realistic person training
Fyou can also do this
F1st fine tune generate pick good ones
Fand fine tune again
Fwell the combination of all parameters make a good workflow
Fso you should research good params there with testing all combinations
EYeah, I'll first try to finetune with only 8 images, will try to find the best # of steps sweet spot until not too much overfit and then generate new images and finetune again
T@Furkan Gözükara SECourses I am not 'urging' u to make a video now but I would like to know if Kohya can be used to train HiDream model (with my personal pics) like Flux ? Can I do it in Massed Compute ?
EIf I train model A for 3200 steps and model B for 5000 steps, both having the same base model, would the result of generating an image be the same for both if for model B I use a checkpoint that's near 3200 steps like model A (save every x epoch/steps)?
And if I train a model for 3200 steps and then retrain it for another 1800 steps for a total of 5000 steps, will that model be as efficient as if I trained a model to 5000 steps directly?
Thank you
And if I train a model for 3200 steps and then retrain it for another 1800 steps for a total of 5000 steps, will that model be as efficient as if I trained a model to 5000 steps directly?
Thank you
Fthat is also on my list but so many stuff 
Fyes it should be same 3200 + 1800 = 5000 directly
Fsince our learning rates are all constant
E@Furkan Gözükara SECourses Is the kohya dreambooth sdxl 48gb config usable with the kohya sdxl lora extract configuration that you have in one of your patreon posts (not sure if I can post image here)?
I am asking because if I leave the default settings of Clamp Quantile at 0.99 and Minimum difference to 0.01, I think it says the text encoder does not need to be extracted or something similar and when I try to use the lora, it simply does not show anything of my trained model compared to if I use the settings in your image of 0.1 for Clamp Quantile and 0.00001 for Minimum difference where now it works fine.
It just made me wonder if that screenshot of the kohya sdxl lora extract settings was up to date with the latest 48gb config of kohya dreambooth settings!
I did see a recent post for lora extract, but it was for flux so I wasn't sure
Thank you for the great support!
I am asking because if I leave the default settings of Clamp Quantile at 0.99 and Minimum difference to 0.01, I think it says the text encoder does not need to be extracted or something similar and when I try to use the lora, it simply does not show anything of my trained model compared to if I use the settings in your image of 0.1 for Clamp Quantile and 0.00001 for Minimum difference where now it works fine.
It just made me wonder if that screenshot of the kohya sdxl lora extract settings was up to date with the latest 48gb config of kohya dreambooth settings!
I did see a recent post for lora extract, but it was for flux so I wasn't sure
Thank you for the great support!
Fmake it 0.000001 to extract
Falso flux post is updated too but flux extract not very good
JDoctor, could you please tell me the best settings for LoRA training for SDXL? I saw your previous post about fine tuning... Have you already posted about LoRA training?
Fwe have dreambooth
Fi recommend after that extract lora
JI see. I'll give it a try. Thank you as always.