Software Engineering Courses (SECourses)•8mo ago

Honestly I'm always wondering how to know the number of epochs and such. Should we just aim for a ce

Honestly I'm always wondering how to know the number of epochs and such. Should we just aim for a certain # of total steps and play with the # of epochs and repeats?

For example, with 15 images 200-300 epochs but how many images repeats? Or is it better 1 epoch and tons of repeats?

EscanorOP•5/16/25, 1:50 PM

I also wonder if with like 15 portrait images we should be using captions or not

EscanorOP•5/16/25, 1:58 PM

Also wondering because the sdxl kohya 48gb config uses 1 epoch

skyrrr•5/16/25, 4:45 PM

Hey, for flux full finetuning is there a solution faster than L40S for the 48gb config file ? Thanks

EEscanor Honestly I'm always wondering how to know the number of epochs and such. Should ...

Furkan Gözükara SECourses•5/16/25, 5:30 PM

repeat is always 1 for flux

Furkan Gözükara SECourses•5/16/25, 5:30 PM

we had to use repeat for sdxl to use more reg images due to how kohya bad coded reg images logic

EEscanor Also wondering because the sdxl kohya 48gb config uses 1 epoch 🤔

Furkan Gözükara SECourses•5/16/25, 5:31 PM

because with sdxl we want to use more reg images

Furkan Gözükara SECourses•5/16/25, 5:31 PM

and only way is more repeat with kohya

Sskyrrr Hey, for flux full finetuning is there a solution faster than L40S for the 48gb ...

Furkan Gözükara SECourses•5/16/25, 5:32 PM

what do you mean?_

EscanorOP•5/16/25, 5:42 PM

Ahh I see. Then for sdxl is it better to use onetrainer instead in your opinion? Considering I have only 8 portrait images and the 5000 reg images of your patreon. Also unsure if I should use captions or not

Thank you

FFurkan Gözükara SECourses what do you mean?_

skyrrr•5/16/25, 6:40 PM

I mean would an H100 run the training significantly faster than the L40S ? I'm currently at 4.8sec/iteration

EEscanor Ahh I see. Then for sdxl is it better to use onetrainer instead in your opinion?...

Furkan Gözükara SECourses•5/16/25, 9:12 PM

yes onetrainer also has EMA

Furkan Gözükara SECourses•5/16/25, 9:12 PM

which works good for SDXL

Sskyrrr I mean would an H100 run the training significantly faster than the L40S ? I'm c...

Furkan Gözükara SECourses•5/16/25, 9:12 PM

yes it is faster

Furkan Gözükara SECourses•5/16/25, 9:12 PM

are you batch size 7?

Furkan Gözükara SECourses•5/16/25, 9:13 PM

4.8 sec is low for H100

EscanorOP•5/16/25, 9:23 PM

Ok for onetrainer, will try.

Is FLUX really that much better than SDXL though even for just 8 training images? Wondering because it's so slow to run on my m2 macbook pro..

EEscanor Ok for onetrainer, will try. Is FLUX really that much better than SDXL though ...

Furkan Gözükara SECourses•5/16/25, 9:24 PM

very very better

Furkan Gözükara SECourses•5/16/25, 9:25 PM

so much difference

EscanorOP•5/16/25, 9:27 PM

Urgh ok. Because on mac I need to use a weird flux model or else tons of errors. Model is flux1-dev-Q8_0.gguf

EEscanor Urgh ok. Because on mac I need to use a weird flux model or else tons of errors....

Furkan Gözükara SECourses•5/16/25, 9:40 PM

i would use mac if only i was developing an app for iphone :d

EscanorOP•5/16/25, 9:41 PM

Meh, for developing from backend to frontend it's great, ai stuff it's shit lol

EscanorOP•5/16/25, 9:41 PM

Compatibility is atrocious

EscanorOP•5/16/25, 9:42 PM

Especially with arm/sillicon chips

EEscanor Especially with arm/sillicon chips

Furkan Gözükara SECourses•5/16/25, 9:54 PM

yep that is why

EscanorOP•5/17/25, 12:04 AM

What is the best workflow for preparing images?

Let's say I have portrait images that are mediocre quality and not 1024x1024. Which of the following is the best workflow:

Workflow A: joycaption,supir 5x upscale with caption,outpaint image sides,downscale to 1024x1024
Workflow B: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with transparent sides
Workflow C: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with black or white sides

EscanorOP•5/17/25, 12:04 AM

Also, I'm unsure which caption type + length I should use with joycaption

EscanorOP•5/17/25, 12:06 AM

And for having square images, for my current dataset, I had used Fooocus to outpaint the sides since the original images were portrait, but would it be acceptable or good to just resize to 1024x1024 and have the sides be black,white or transparent..? Or will that confuse the training?

EscanorOP•5/17/25, 12:06 AM

Sorry for many questions, thank you very much for the help!

FFurkan Gözükara SECourses are you batch size 7?

skyrrr•5/17/25, 8:25 AM

No I mean I'm currently using the 48go config with an L40S and getting around4.8 sec. I have a deadline and would need the training to be twice faster, so I wondered if that was achievable with an H100 SXM. Sorry if I explained myself wrong

EEscanor What is the best workflow for preparing images? Let's say I have portrait image...

Furkan Gözükara SECourses•5/17/25, 10:59 AM

for flux i dont use captions i dont recommend

Sskyrrr No I mean I'm currently using the 48go config with an L40S and getting around4.8...

Furkan Gözükara SECourses•5/17/25, 11:00 AM

ah yes go with h100 it will be faster directly

FFurkan Gözükara SECourses for flux i dont use captions i dont recommend 😄

EscanorOP•5/17/25, 1:21 PM

Well it wasn't for flux specifically, it was to have captions from image to upscale them with SUPIR!

What I'm unsure of is the best workflow to prepare a dataset of mediocre quality to increase its quality...

And also if I can just resize to 1024x1024 a portrait images and just leave the sides transparent

EEscanor Well it wasn't for flux specifically, it was to have captions from image to upsc...

Furkan Gözükara SECourses•5/17/25, 3:15 PM

for supir caption matters

Furkan Gözükara SECourses•5/17/25, 3:15 PM

you can use joy caption

EscanorOP•5/17/25, 3:17 PM

Yeah I am, just wondering if it's better to use "descriptive" caption type or "training" and the caption length

Also hesitant for more specific words. For example, my photo has "snake bites" piercings which is essentially a ring on each side of the bottom lip, but I'm unsure if I should put "snake bites ring piercings" or just like "two lip ring piercings"

EscanorOP•5/17/25, 3:18 PM

I'm literally trying it out right now on a h100 on massed compute right now

EEscanor Yeah I am, just wondering if it's better to use "descriptive" caption type or "t...

Furkan Gözükara SECourses•5/17/25, 3:18 PM

no idea you better test

FFurkan Gözükara SECourses no idea you better test

EscanorOP•5/17/25, 3:20 PM

Ok, thank you

Only other thing I'm wondering is this: if I have a portrait image, can I just resize to 1024x1024 images and just leave the sides transparent or white/black or do I absolutely have to outpaint the sides if I want to use this image for training in my dataset?

Thanks!

EEscanor Ok, thank you 🙂 Only other thing I'm wondering is this: if I have a portrait ...

Furkan Gözükara SECourses•5/17/25, 3:40 PM

you can leave empty transparent

EscanorOP•5/17/25, 3:54 PM

oh really?? Wow.. This will make things so much easier...

I had spent to much time trying to outpaint something decent for the sides with Fooocus..

Damnit lol, glad I asked thank you very much

EEscanor oh really?? Wow.. This will make things so much easier... 🤦 I had spent to muc...

Furkan Gözükara SECourses•5/17/25, 3:57 PM

you are welcome

TheRealDude•5/17/25, 9:49 PM

TTheRealDude Hi

Furkan Gözükara SECourses•5/17/25, 10:36 PM

mlado•5/18/25, 9:48 AM

Hello everyone

I am new to the AI world, started learning around a week ago, and my goal is to create a realistic-looking Instagram influencer with the same face and body (body seems hard but at least face hah). I have a dataset of 70 images, which are all 1024x1024 and good quality I think, and I will be training a lora in Fluxgym because it seems beginner friendly.
I just wanted to ask for your suggestions and advice on a few things. What would you recommend in terms of trains per image, number of epochs, and LoRA rank for this kind of realistic, identity-focused project?

(also thank you very much for all your research and insights @Furkan Gözükara SECourses)

Mmlado Hello everyone 👋 I am new to the AI world, started learning around a week ago, ...

Timson•5/18/25, 10:07 AM

start (and end) with fine tuning, skip LORA step. It's not much harder for beginner, basically same process but drastically better in end result.
If you want to focus on the very best result for body + face train 2 separate models on 2 separate datasets - one focused on body (full, medium shots, poses) the other on face (mid, close shots), then use 2 step generation with face model over image generated on body model.
If you want easier process - balance your dataset well for close and mid shots (30% mid, 70% close)

mlado•5/18/25, 10:08 AM

Okay thank you, I have a 4070 ti super, is that sufficient for fine tuning?

mlado•5/18/25, 10:09 AM

And yeah tbh I watched doctor futkans video about fine tuning vs lora and he also said that lora is inferior

Timson•5/18/25, 10:09 AM

Idaely you would have to rent A6000, it's 0.3$ per hour and training would take 6-10 hours

mlado•5/18/25, 10:10 AM

Okay, thats through runpod right

Timson•5/18/25, 10:10 AM

no, massedcompute

Honestly I'm always wondering how to know the number of epochs and such. Should we just aim for a ce

Similar Threads