Software Engineering Courses (SECourses)•6mo ago

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

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?

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

Furkan Gözükara SECourses•7/10/25, 11:06 AM

100%

Furkan Gözükara SECourses•7/10/25, 11:06 AM

pay attention to your empty vram

Furkan Gözükara SECourses•7/10/25, 11:06 AM

it uses 23 gb

Furkan Gözükara SECourses•7/10/25, 11:06 AM

you can increase block swap count if you cant have 23 gb free due to other apps

FFurkan Gözükara SECourses pay attention to your empty vram

Elli_OakenOP•7/10/25, 2:00 PM

Ok, could I do say 6 different colors of the same bag (same shape) in one checkpoint? I guess I would need captioning then.

EElli_Oaken Ok, could I do say 6 different colors of the same bag (same shape) in one checkp...

Furkan Gözükara SECourses•7/10/25, 2:39 PM

i think you can train different colors of same bag

underdog•7/10/25, 11:15 PM

Hello,

I'm trying to run grid generator but I get this error, do you know why?
I'm familiar with comfyui but this is my first time trying swarmui

underdog•7/10/25, 11:16 PM

Uunderdog Click to see attachment

Furkan Gözükara SECourses•7/11/25, 12:05 AM

hard to know without seeing all

Furkan Gözükara SECourses•7/11/25, 12:05 AM

you can post debug logs here

underdog•7/11/25, 12:06 AM

I was able to get it working. I pulled the latest code and it started working.

underdog•7/11/25, 1:35 AM

while generating images in a grid with multiple models, how does inpainting work? which image does it take as reference?

Uunderdog while generating images in a grid with multiple models, how does inpainting wor...

Furkan Gözükara SECourses•7/11/25, 1:37 AM

what kind of inpainting?

Furkan Gözükara SECourses•7/11/25, 1:37 AM

segment?

underdog•7/11/25, 1:37 AM

yes, to make the face look more like the original character and to make it more clear

Uunderdog yes, to make the face look more like the original character and to make it more ...

Furkan Gözükara SECourses•7/11/25, 1:40 AM

it inpaints every image

Furkan Gözükara SECourses•7/11/25, 1:40 AM

it generates

Furkan Gözükara SECourses•7/11/25, 1:40 AM

it is part of prompt

underdog•7/11/25, 1:43 AM

yes that I understand. While generating one image with one model it takes the first image that we generated as reference and inpaints the face right? but in grid generation, I didnt have a first image and I directly started generating images. My doubt is what reference would it take for inpainting if we dont have our first image generated?

Uunderdog yes that I understand. While generating one image with one model it takes the fi...

Furkan Gözükara SECourses•7/11/25, 1:47 AM

it takes every generated image

Furkan Gözükara SECourses•7/11/25, 1:47 AM

and then inpaint face

underdog•7/11/25, 1:50 AM

Oh okay. I thought "1" in 1, 0.7, 0.5 meant take the first image as reference for inpainting the face <segment:yolo-face_yolov9c.pt-1,0.7,0.5//cid=11> that is why I was confused and wanted to know what would happen for the first image generation if we didn't have our first image generated.

Uunderdog Oh okay. I thought "1" in 1, 0.7, 0.5 meant take the first image as reference fo...

Furkan Gözükara SECourses•7/11/25, 1:55 AM

1 means first face in image

underdog•7/11/25, 1:57 AM

I completely misunderstood. Thanks for the clarification.

underdog•7/11/25, 2:06 AM

can we use ipadapter to use a reference image/face so the generated images are consistent?
similar to segment, do we have anything for ipadapter like this - <ipadapter:test_image.jpg:0.5> ?

Uunderdog can we use ipadapter to use a reference image/face so the generated images are c...

Furkan Gözükara SECourses•7/11/25, 9:09 AM

we have this but there could be other ways too

Furkan Gözükara SECourses•7/11/25, 9:09 AM

https://www.patreon.com/posts/118469722

Patreon

Ultra Advanced InstantID Gradio APP - Automatic Installers for Wind...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

underdog•7/11/25, 9:09 AM

Thank you

I will take a look!

underdog•7/11/25, 9:05 PM

I noticed you are using adafactor, what is your take on prodigy? the learning rate will adapt accordingly right? it might consume more vram but what do you think the result quality would be with prodigy?

I might be wrong but because the learning rate adapts, I think we might eliminate the risk of overfitting. Your advise would be helpful.

Uunderdog I noticed you are using adafactor, what is your take on prodigy? the learning ra...

Furkan Gözükara SECourses•7/11/25, 9:14 PM

i tested them and none of the others yielded as realistic as adafactor

Furkan Gözükara SECourses•7/11/25, 9:14 PM

i did a big test recently

underdog•7/11/25, 9:15 PM

Thank you

for 15 image dataset, were you able to figure what the unet learning rate was ideal?

Uunderdog Thank you 🙂 for 15 image dataset, were you able to figure what the unet learnin...

Furkan Gözükara SECourses•7/11/25, 9:15 PM

yes our rate is ideal

Furkan Gözükara SECourses•7/11/25, 9:15 PM

just do at least 200 epoch

Furkan Gözükara SECourses•7/11/25, 9:15 PM

but you can do up to like 3000 and compare

FFurkan Gözükara SECourses i did a big test recently

underdog•7/11/25, 9:15 PM

are you able to share the results for learning purposes?

underdog•7/11/25, 9:21 PM

also for batch size, what number do you recommend? I'm renting A100 gpu. I'm testing with a batch size of 1 for 15 images. I want to increase it but I'm afraid the model might overfit. I can test them myself but I'm paying by the hour so testing all of them is going to burn me out

Uunderdog also for batch size, what number do you recommend? I'm renting A100 gpu. I'm tes...

Furkan Gözükara SECourses•7/11/25, 9:22 PM

i recommend bs 1

Uunderdog are you able to share the results for learning purposes?

Furkan Gözükara SECourses•7/11/25, 9:23 PM

i planned the test but later didnt have time

underdog•7/11/25, 9:24 PM

Thank you very much. Your testing evaluation is saving my time

underdog•7/11/25, 9:28 PM

one more question
you used constant LR scheduler, why not cosine or cosine with restarts?

With the constant scheduler, the model doesn't learn. The training loss doesn't decrease. Is this okay?

Uunderdog one more question you used constant LR scheduler, why not cosine or cosine with ...

Furkan Gözükara SECourses•7/11/25, 9:33 PM

training loss useless

Furkan Gözükara SECourses•7/11/25, 9:33 PM

constant also gives you chance to continue perfectly fine

Konstantin•7/12/25, 5:59 PM

What could be the problem, I tried the kohya-ss/sd-scripts branch for lumina 2 (feature-lumina-image) for lumina_train.py. There were a couple of typos img instead of noisy_model_input, but after correcting them, the training started, the loss even goes down, but at the first generation of samples I get noise. If I add "sample_at_first = true" then I see that before training the image is generated normally, but from the very beginning of training there is noise. Now I have
max_timestep = 1000
guidance_scale = 4
timestep_sampling = "nextdit_shift"
discrete_flow_shift = 6.0
model_prediction_type = "raw"
optimizer_args = [ "weight_decay=0.00",]
optimizer_type = "PagedLion8bit"

I tried changing the optimizer, timestep_sampling, discrete_flow_shift, all the same, no results... Already at the 30th step of training there is noise.

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

Similar Threads