Software Engineering Courses (SECourses)•9mo ago

Based off the example 256 img training log for the 48GB-batch size 1 config, should I expect to trai

Based off the example 256 img training log for the 48GB-batch size 1 config, should I expect to train until ~50K steps before seeing decent results?

HHoot Based off the example 256 img training log for the 48GB-batch size 1 config, sho...

Furkan Gözükara SECourses•4/23/25, 12:48 PM

50k probably overkill

Furkan Gözükara SECourses•4/23/25, 12:48 PM

i had best checkpoint at 70 epoch

Furkan Gözükara SECourses•4/23/25, 12:48 PM

17920 steps

HootOP•4/23/25, 12:54 PM

what epoch was your best checkpoint for the batch 7 config? I have 252 imgs so I’ve been training until 111 epochs to reach 4K steps but never went higher than that.

HHoot 🫣 what epoch was your best checkpoint for the batch 7 config? I have 252 imgs s...

Furkan Gözükara SECourses•4/23/25, 12:56 PM

i dont remember but probably similar

Furkan Gözükara SECourses•4/23/25, 12:56 PM

you know epoch doesnt change but step count changes

Furkan Gözükara SECourses•4/23/25, 12:56 PM

with batch size

HootOP•4/23/25, 1:03 PM

So the higher the batch size, the lower amount of steps you’ll need to achieve a quality result?

HootOP•4/23/25, 1:03 PM

Like if I rented an H200 and set the batch count to 14, I would likely need fewer than 4K steps to get the same quality as batch size 7?

HHoot So the higher the batch size, the lower amount of steps you’ll need to achieve a...

Furkan Gözükara SECourses•4/23/25, 1:05 PM

batch size meaning is at 1 step it process more images

Furkan Gözükara SECourses•4/23/25, 1:06 PM

so we are still processing like equal number of samples (the model trained number of images)

Furkan Gözükara SECourses•4/23/25, 1:06 PM

but the model weights are updated lesser

Furkan Gözükara SECourses•4/23/25, 1:06 PM

since batch size changes averaged

HootOP•4/23/25, 1:07 PM

Gotcha, thank you for clarifying!

yemmlie•4/27/25, 3:59 PM

hey just joined to try get better settings for kohya finetune of flux and see discord pop up so thought I'd ask here. 48gb RTX 6000 Ada should be enough to finetune right? going mad here keeps getting CUDA out of memory even with 512x512, batch 1, fp16, adam8, gradient_checkpointing and everything i can find to reduce memory, and feel i must be doing something wrong

its really odd

yemmlie•4/27/25, 4:46 PM

nm i used one of the scripts from the patreon for 48gb and it works - god only knows which setting it is

Thanks! much worth the patreon sub!

ReadyToRetireLB_562•4/28/25, 9:10 PM

@Furkan Gözükara SECourses The onetrainer sdxl config files load into fine tune, not lora. is this the intent?

RReadyToRetireLB_562 <@205854764540362752> The onetrainer sdxl config files load into fine tune, n...

Furkan Gözükara SECourses•4/28/25, 9:17 PM

yep i dont have lora config sadly

Yyemmlie nm i used one of the scripts from the patreon for 48gb and it works - god only k...

Furkan Gözükara SECourses•4/28/25, 9:17 PM

great

Punceka•4/29/25, 5:23 PM

Hi, I've been trying to merge 2 Flux LoRas via Kohya SS together and I keep failing.
Say, if normally I set these 2 Loras to the strenght of 1 and 0.6, what ratios should I put in there? And then, sometimes it merges a lora to the size of 24Gb or 40Gb+ if I switch the save precision so I'm not really sure what settings in general there would be the best

cat cafe•4/29/25, 11:57 PM

Is ohwx an especially good instance prompt or would using any another string of rare special characters would be just as good?

Ccat cafe Is ohwx an especially good instance prompt or would using any another string of ...

Furkan Gözükara SECourses•4/30/25, 12:50 AM

ohwx good

Furkan Gözükara SECourses•4/30/25, 12:50 AM

rare token

RyanBelnap•4/30/25, 9:05 PM

Just got around to dreambooth videos and I gotta say this is amazing we can do this. You are a god send my friend...Keep up the great work!

asap17z•4/30/25, 9:12 PM

Is it a bad idea to use regularisation dataset with a cartoon dataset for fine tuning?

RRyanBelnap Just got around to dreambooth videos and I gotta say this is amazing we can do t...

Furkan Gözükara SECourses•4/30/25, 9:21 PM

ty so much

Aasap17z Is it a bad idea to use regularisation dataset with a cartoon dataset for fine t...

Furkan Gözükara SECourses•4/30/25, 9:21 PM

it depends. for flux we dont use. for sdxl if your reg images are also cartoon good

lerois1•5/1/25, 9:45 PM

I admire your work. As a Patreon member and Massed Compute user ($100+ spent), I'm experiencing an issue with my Dreambooth training.

My training with 277 properly prepared images, using background filling, is progressing using the "Fine Tuning / DreamBooth: 256 Training Images & Batch Size is 1 : Best Epoch is 70" setting. This should finish within the hour on Massed Compute, totaling 277 x 70 steps. I have checkpoint intervals every 10 epochs and have six models so far.

The first checkpoint showed excellent results, but subsequent epochs have produced increasingly blurry images.

Would it be possible to discuss this with you, perhaps via Skype? I would appreciate any guidance you can offer, either freely or through a specialized support option.

Llerois1 I admire your work. As a Patreon member and Massed Compute user ($100+ spent), I...

Furkan Gözükara SECourses•5/1/25, 10:59 PM

thank you so much and welcome. replied you from email

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/2/25, 7:41 PM

can you explain this, why are the eyes so bad? did sdxl lora training, 30 train pics, all great quality, many close ups on face, 150 epochs, repeat 1, 1024 resolution, biglovexl2 as base model.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/2/25, 7:41 PM

training on civitai

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/2/25, 7:47 PM

training-data, i see network rank 32. don't know if its possible to change that on training

train-data.txt1.41KB

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭can you explain this, why are the eyes so bad? did sdxl lora training, 30 train ...

Furkan Gözükara SECourses•5/3/25, 12:44 AM

are you using our config_?

Furkan Gözükara SECourses•5/3/25, 12:44 AM

which config name?

Furkan Gözükara SECourses•5/3/25, 12:44 AM

also some people reporting bad results

Furkan Gözükara SECourses•5/3/25, 12:44 AM

i am gonna train now to see

FFurkan Gözükara SECourses are you using our config_?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/3/25, 6:30 AM

this was civitai training, I used what you said about ohwx woman token, 150 epochs, repeat 1.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/3/25, 6:33 AM

where can we download flux dev fast now? black labs has it gated, so can't use the download link in script...

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭this was civitai training, I used what you said about ohwx woman token, 150 epoc...

Furkan Gözükara SECourses•5/3/25, 12:05 PM

ah i dont know

Furkan Gözükara SECourses•5/3/25, 12:05 PM

i trained on my local pc yesterday

Furkan Gözükara SECourses•5/3/25, 12:05 PM

our config

Furkan Gözükara SECourses•5/3/25, 12:05 PM

latest kohya

Furkan Gözükara SECourses•5/3/25, 12:05 PM

perfect results

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/3/25, 12:34 PM

Oh awesome, you recommend kohya for sdxl lora and checkpoint training now?

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭Oh awesome, you recommend kohya for sdxl lora and checkpoint training now?

Furkan Gözükara SECourses•5/3/25, 12:37 PM

i tested flux

Furkan Gözükara SECourses•5/3/25, 12:37 PM

but we have sdxl dreambooth config too

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/3/25, 12:37 PM

Ok, so onetrainer gives slightly better results for sdxlin your opinion?

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭Ok, so onetrainer gives slightly better results for sdxlin your opinion?

Furkan Gözükara SECourses•5/3/25, 12:52 PM

yep

asap17z•5/3/25, 4:00 PM

Hey Doc, how you doing today, I was just wondering if it’s possible to create a stylised flux model in place of a stylised lora

FFurkan Gözükara SECourses thank you so much and welcome. replied you from email

lerois1•5/3/25, 4:11 PM

I am pleased to inform you that the dreambooth training has produced excellent results. It seems the previous difficulty arose from an incorrect learning rate.
I have encountered a remaining challenge: all generated images tend to resemble the model, irrespective of the prompt utilized. Although the model was trained using the prompt "ohwx" and the class "woman," this issue persists even when the prompt is omitted or when a different subject, such as "man," is specified.
Could you please offer some guidance on the potential cause of this occurrence? My objective is to generate images of other women/man in certain scenes who do not precisely replicate the trained model.

Based off the example 256 img training log for the 48GB-batch size 1 config, should I expect to trai

Similar Threads