Software Engineering Courses (SECourses)•6mo ago

HI! I have a question about lora-training and captioning. A friend of mine suggest that when you cap

HI! I have a question about lora-training and captioning. A friend of mine suggest that when you caption your object your not supposed to mention the object in the image at all, only everything around in like a semantic mask and where the object-tag works as the caption for the lora-object. On youtube-tutorials on the other hand everyone I've seen says that you should in detail concistently describe the object and only mention shortly the rest of the image.

What would you guys say is the case here?

Uunderdog update -Dwayne johnson training works fine. I get good results around the 3000th...

Furkan Gözükara SECourses•7/9/25, 1:54 PM

sadly not

Furkan Gözükara SECourses•7/9/25, 1:54 PM

you need to evaluate yourself

Uunderdog Another question - In the dataset Dwayne johnson does not have hair but when I g...

Furkan Gözükara SECourses•7/9/25, 1:54 PM

yes it is about prompt

Furkan Gözükara SECourses•7/9/25, 1:54 PM

just add bald :d

EElli_Oaken HI! I have a question about lora-training and captioning. A friend of mine sugge...

Furkan Gözükara SECourses•7/9/25, 1:54 PM

captioning is needed only when you do a general fine tuning

Furkan Gözükara SECourses•7/9/25, 1:54 PM

ohwx man , ohwx class works best

Furkan Gözükara SECourses•7/9/25, 1:55 PM

then after training in details give your prompt

Furkan Gözükara SECourses•7/9/25, 1:55 PM

detailed captioning only reduces your trained subject resemblance

Furkan Gözükara SECourses•7/9/25, 1:55 PM

because even if you caption the elements you dont want to be learnt, it still leaks info of what you want to be learnt

FFurkan Gözükara SECourses detailed captioning only reduces your trained subject resemblance

Elli_OakenOP•7/9/25, 2:03 PM

Ok, so no captions only caption tag in kohya? Should I use a high or low network dim for 50 images that are very consistent, I'm training a bag

EElli_Oaken Ok, so no captions only caption tag in kohya? Should I use a high or low network...

Furkan Gözükara SECourses•7/9/25, 2:09 PM

i recommend use my config and dont change

Furkan Gözükara SECourses•7/9/25, 2:09 PM

you can use folder name as caption

Furkan Gözükara SECourses•7/9/25, 2:09 PM

so like

Furkan Gözükara SECourses•7/9/25, 2:09 PM

1_ohwx man

Furkan Gözükara SECourses•7/9/25, 2:09 PM

or have imagefile.txt files and write each one of them like ohwx man

Furkan Gözükara SECourses•7/9/25, 2:09 PM

both works

FFurkan Gözükara SECourses you can use folder name as caption

Elli_OakenOP•7/9/25, 8:06 PM

Ex 24GB_GPU_Quality_Tier_1_22950MB_14.1_Second_IT_Trains_T5_and_T5_Attention.json ?

EElli_Oaken Ex 24GB_GPU_Quality_Tier_1_22950MB_14.1_Second_IT_Trains_T5_and_T5_Attention.jso...

Furkan Gözükara SECourses•7/9/25, 9:22 PM

that is config

Furkan Gözükara SECourses•7/9/25, 9:22 PM

folder name means

Furkan Gözükara SECourses•7/9/25, 9:23 PM

Jamiewhelton•7/9/25, 10:14 PM

does this matter? the outdated version

JJamiewhelton does this matter? the outdated version

Furkan Gözükara SECourses•7/9/25, 10:20 PM

nope

Furkan Gözükara SECourses•7/9/25, 10:20 PM

folder name doesnt matter we use latest version

FFurkan Gözükara SECourses yes it is about prompt

underdog•7/10/25, 1:34 AM

Thank you

The finetuned checkpoint is about 23gb, is there a way to reduce the size without losing the quality?

Uunderdog Thank you 🙂 The finetuned checkpoint is about 23gb, is there a way to reduce th...

Furkan Gözükara SECourses•7/10/25, 1:35 AM

you can do lora

Furkan Gözükara SECourses•7/10/25, 1:35 AM

i dont remember if supporting save as fp8

FFurkan Gözükara SECourses you can do lora

underdog•7/10/25, 1:36 AM

I tried lora but I like the outputs that im getting from finetuning. I see fewer overfits and more consistency in finetuned model.

Uunderdog I tried lora but I like the outputs that im getting from finetuning. I see fewer...

Furkan Gözükara SECourses•7/10/25, 1:45 AM

ye fine tune best

Elli_OakenOP•7/10/25, 8:18 AM

I'm sonfused and have to ask again,

am I wrong in that you still can't train a full checkpoint below 80 gb vram? And finetuning means full checkpoint right?

FFurkan Gözükara SECourses folder name means

Elli_OakenOP•7/10/25, 10:00 AM

The lora turned out really bad. No shape from prodct at all, it only makes like a genral bag like it dosn't under geometry, or the shape is to complex

EElli_Oaken I'm sonfused and have to ask again, am I wrong in that you still can't train a...

Furkan Gözükara SECourses•7/10/25, 10:45 AM

full checkpoint can be trained as low as 6 gb vram

Furkan Gözükara SECourses•7/10/25, 10:45 AM

6GB_GPU_5400MB_14.1_second_it_Tier_3_512px.json

Furkan Gözükara SECourses•7/10/25, 10:45 AM

this is same as 48 gb config quality

Furkan Gözükara SECourses•7/10/25, 10:45 AM

8GB_GPU_7500MB_20.5_second_it_Tier_1.json

FFurkan Gözükara SECourses 6GB_GPU_5400MB_14.1_second_it_Tier_3_512px.json

Elli_OakenOP•7/10/25, 10:55 AM

Could i use the 24 gb config for my rtx 4090?