Software Engineering Courses (SECourses)•5mo ago

I played a bit with Wan 2.1 Lora training too, it seemed easier or more forgiving. But only tested i

I played a bit with Wan 2.1 Lora training too, it seemed easier or more forgiving. But only tested it

JJonkoXL I played a bit with Wan 2.1 Lora training too, it seemed easier or more forgivin...

cyberbol•8/15/25, 3:11 PM

For wan2.1 I did some Loras

https://tensor.art/u/842260292657429385/models

CyberAImania's Models

JonkoXLOP•8/15/25, 3:15 PM

OOh nice :p

JonkoXLOP•8/15/25, 3:15 PM

what's your learning_rate and how many steps did you train for?

cyberbol•8/15/25, 3:18 PM

Wan2.1? I don't remember..but I Keep on drive all dataset and configs.

Ccyberbol 4090 (24 GB VRAM) – it was fine, but I’m now collecting a better dataset for my ...

Sakami•8/15/25, 3:19 PM

thanks - right now im doing Wan2.2 training on my dataset ll see how it turns out

SSakami thanks - right now im doing Wan2.2 training on my dataset ll see how it turns ou...

cyberbol•8/15/25, 3:20 PM

Let me know please. Share also what datasets you used if you can. Just photo's or mixed or video clips?

Sakami•8/15/25, 3:22 PM

due to nda i cant share dataset but these are studio photos and few natural shots of a person. Unfortunetly only photos for now

masterblaster3606•8/15/25, 5:28 PM

hello, i got out of memory error using massed-compute 2x RTX 6000 ADA

Batch_Size_7_48GB_GPU_46250MB_29.1_second_it_Tier_1_multi-GPU.json5.03KB

Mmasterblaster3606 hello, i got out of memory error using massed-compute 2x RTX 6000 ADA

JonkoXLOP•8/15/25, 5:34 PM

training what?

JJonkoXL training what?

masterblaster3606•8/15/25, 6:04 PM

40 images of a human for Flux

Sakami•8/15/25, 6:07 PM

No LoRA vs LoRA Wan2.2

Sakami•8/15/25, 6:08 PM

she is little bit damaged tho its 40 epoch [65img's dataset] 32DIM

JonkoXLOP•8/15/25, 7:49 PM

learning rate?

Mmasterblaster3606 hello, i got out of memory error using massed-compute 2x RTX 6000 ADA

Furkan Gözükara SECourses•8/15/25, 8:00 PM

did you load lora config into dreambooth or dreambooth into lora?

Furkan Gözükara SECourses•8/15/25, 8:00 PM

config name?

FFurkan Gözükara SECourses did you load lora config into dreambooth or dreambooth into lora?

masterblaster3606•8/15/25, 8:07 PM

No, I loaded dreambooth config into Dreambooth. I tried both 48GB_GPU_28200MB_6.3_second_it_Tier_1 and Batch_Size_7_48GB_GPU_46250MB_29.1_second_it_Tier_1. out of memory errors for both

Mmasterblaster3606 No, I loaded dreambooth config into Dreambooth. I tried both 48GB_GPU_28200MB_6....

Furkan Gözükara SECourses•8/15/25, 8:25 PM

what is your images resolution?

Furkan Gözükara SECourses•8/15/25, 8:25 PM

i recommend restart machine and use this config load into dreambooth

Furkan Gözükara SECourses•8/15/25, 8:25 PM

will work 100%

Furkan Gözükara SECourses•8/15/25, 8:25 PM

48GB_GPU_28200MB_6.3_second_it_Tier_1.json

JJonkoXL learning rate?

Sakami•8/15/25, 8:52 PM

3e-4

FFurkan Gözükara SECourses what is your images resolution?

masterblaster3606•8/15/25, 8:53 PM

1024x1024. It works in single GPU, but always memory error when I try to use dual GPU

SSakami 3e-4

JonkoXLOP•8/15/25, 8:54 PM

So far I tried 2e-4 and 3e-4 and it seems 3e-4 overburned my Lora and 2e-4 is just a little bit underpowered (need ~1.2x lora strength)

JonkoXLOP•8/15/25, 8:54 PM

I will try 0.000234 next as learning rate

JonkoXLOP•8/15/25, 8:54 PM

just a little bit more than 2e-4

JonkoXLOP•8/15/25, 8:55 PM

At ~ 1000 steps

JonkoXLOP•8/15/25, 8:56 PM

I have no idea if amount of images affects the learning rate, if we keep the steps the same

JJonkoXL I have no idea if amount of images affects the learning rate, if we keep the ste...

Sakami•8/15/25, 8:59 PM

maybeeee

JJonkoXL I will try 0.000234 next as learning rate

Sakami•8/15/25, 8:59 PM

let me know

JonkoXLOP•8/15/25, 8:59 PM

Sure thing but I think I am on to something

Mmasterblaster3606 1024x1024. It works in single GPU, but always memory error when I try to use dua...

Furkan Gözükara SECourses•8/15/25, 9:00 PM

dual gpu dreambooth need 96 gb

Furkan Gözükara SECourses•8/15/25, 9:00 PM

or lets say at least 80 gb

Furkan Gözükara SECourses•8/15/25, 9:00 PM

gpus

FFurkan Gözükara SECourses dual gpu dreambooth need 96 gb

masterblaster3606•8/15/25, 9:03 PM

i got memory errors on 2x RTX A6000 (48GB)

Mmasterblaster3606 i got memory errors on 2x RTX A6000 (48GB)

Furkan Gözükara SECourses•8/15/25, 9:04 PM

yes you cant train dual gpu dreambooth

Furkan Gözükara SECourses•8/15/25, 9:04 PM

but you can train lora

FFurkan Gözükara SECourses yes you cant train dual gpu dreambooth

masterblaster3606•8/15/25, 9:07 PM

i see in your training log 256_img_BS_7_Fine_Tuning.txt that you trained 8x GPU dreambooth with NVIDIA RTX A6000 VRAM 48577MB Arch

Mmasterblaster3606 i see in your training log 256_img_BS_7_Fine_Tuning.txt that you trained 8x GPU ...

Furkan Gözükara SECourses•8/15/25, 9:32 PM

yes but that 8x gpu was used for different trainings

Furkan Gözükara SECourses•8/15/25, 9:32 PM

not as a combination

Furkan Gözükara SECourses•8/15/25, 9:32 PM

lora works with multiple gpu but dreambooth fails unless you get big vram gpus

Jamiewhelton•8/16/25, 8:12 AM

so whats better for fine tuned image gen realism right now?

dev or krea?

pros and cons? can someone give me their observations?

JJamiewhelton so whats better for fine tuned image gen realism right now? dev or krea? pros...

Furkan Gözükara SECourses•8/16/25, 8:13 AM

i posted grids and hopefully will make a video today

Furkan Gözükara SECourses•8/16/25, 8:13 AM

still working on wan 2.2 presets

Jamiewhelton•8/16/25, 8:17 AM

where can I see your grids, patreon?

Jamiewhelton•8/16/25, 8:18 AM

and how about wan 2.2? kling 2.1 master still better?

JJamiewhelton and how about wan 2.2? kling 2.1 master still better?

Furkan Gözükara SECourses•8/16/25, 8:18 AM

i dont use kling sadly

JJamiewhelton and how about wan 2.2? kling 2.1 master still better?

Furkan Gözükara SECourses•8/16/25, 8:18 AM

girds here https://www.patreon.com/posts/112099700

Patreon

Kohya FLUX Fine Tuning (Full Checkpoints) Training Full Tutorial Fo...

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

Furkan Gözükara SECourses•8/16/25, 8:19 AM

Jamiewhelton•8/16/25, 8:22 AM

I appreciate you taking the time to make that

Jamiewhelton•8/16/25, 8:22 AM

I personally don't see any differences/patterns

I played a bit with Wan 2.1 Lora training too, it seemed easier or more forgiving. But only tested i

Similar Threads

Similar Threads

Similar Threads