Software Engineering Courses (SECourses)•16mo ago

Doc @Dr. Furkan Gözükara , I am using default (16 bit) setting for DType on massed compute. When I c

Doc @Dr. Furkan Gözükara , I am using default (16 bit) setting for DType on massed compute. When I come back with the final lora on my humble home PC and FP8 setting, will I see any drop in quality?

Ddruhl Training completed using 2 GPUs on Mass Compute. The experience was great, much ...

Furkan Gözükara SECourses•9/8/24, 3:01 AM

amazing

Xxlads file with .tar are good to use inside of swarmui? I downloaded "phhofm_sisr_mode...

Furkan Gözükara SECourses•9/8/24, 3:01 AM

tar is like zip file i think

druhlOP•9/8/24, 3:01 AM

I ran Rank 1 Slow with T5 attention mask checked. I guess, the wait was worth it 'cuz the quality is amazing.

Ddruhl Doc @Dr. Furkan Gözükara , I am using default (16 bit) setting for DType on mass...

Furkan Gözükara SECourses•9/8/24, 3:01 AM

very very small

Furkan Gözükara SECourses•9/8/24, 3:01 AM

it may

Ddruhl I ran Rank 1 Slow with T5 attention mask checked. I guess, the wait was worth it...

Furkan Gözükara SECourses•9/8/24, 3:01 AM

it improves slightly

Furkan Gözükara SECourses•9/8/24, 3:02 AM

it is accurate approach to do small test first

Furkan Gözükara SECourses•9/8/24, 3:02 AM

man i am gonna make a huge perspective dataset

Furkan Gözükara SECourses•9/8/24, 3:02 AM

flux is just amazing

Furkan Gözükara SECourses•9/8/24, 3:02 AM

0301-photo20of20ohwx20man20as20a20Bedouin20trader20le-flux1-dev-1596466095.png

Furkan Gözükara SECourses•9/8/24, 3:02 AM

when i add perspectives i expect my results to 10x get better

Furkan Gözükara SECourses•9/8/24, 3:03 AM

0302-photo20of20ohwx20man20standing20on20a20cliff20ov-flux1-dev-1596466102.png

xlads•9/8/24, 3:05 AM

Wild!

FFurkan Gözükara SECourses Click to see attachment

druhlOP•9/8/24, 3:06 AM

You gave me an amazing style-training idea with this post

druhlOP•9/8/24, 3:06 AM

When does the style training vid come out, Doc?

druhlOP•9/8/24, 3:06 AM

I can probably set up one next weekend now.

Ddruhl When does the style training vid come out, Doc?

Furkan Gözükara SECourses•9/8/24, 3:09 AM

i dont know yet some people requesting onetrainer video :d

Furkan Gözükara SECourses•9/8/24, 3:09 AM

got to sleep now hopefully see u later

druhlOP•9/8/24, 3:10 AM

Sweet dreams!

druhlOP•9/8/24, 4:03 AM

So, I just terminate the massed compute instance and it all goes away correct?

DDEXperience Solutions GPU 2, 90 C* I think it needs maintenance

druhlOP•9/8/24, 4:20 AM

Dude, you're gonna start a fire

at massed compute

xlads•9/8/24, 5:15 AM

How do i get a Unet folder path inside of swarmui so i can add flux model

druhlOP•9/8/24, 5:16 AM

Its in models > unet

xlads•9/8/24, 5:32 AM

It’s not showing up for me

FFurkan Gözükara SECourses Click to see attachment

AIGambino•9/8/24, 5:37 AM

whats the prompt here? lol love it

protector131090•9/8/24, 7:15 AM

Hello! whats the news on Flux Dreambooth training in Kohya on 24 vram ? possible already?

Xxlads It’s not showing up for me

druhlOP•9/8/24, 7:24 AM

Create it.

Neuromasters•9/8/24, 8:33 AM

are there any other samplers?

Neuromasters•9/8/24, 8:33 AM

it's Flux

Neuromasters•9/8/24, 8:35 AM

Neuromasters•9/8/24, 8:54 AM

And which upskellers are the best if you apply the Hires fix at the moment

magical•9/8/24, 9:05 AM

for generating is there a reason to use the fp8 model instead of the 23.8gb dev model? is it faster to use without loss of quality?

magical•9/8/24, 9:12 AM

btw this link in the kaggle notebook is not working https://huggingface.co/Kijai/flux-fp8/resolve/main/flux1-dev-fp8.safetensors so I'm downloading "flux1-dev-fp8-e4m3fn.safetensors" and " flux1-dev-fp8-e5m2.safetensors" separately from here https://huggingface.co/Kijai/flux-fp8/tree/main

Mmagical btw this link in the kaggle notebook is not working https://huggingface.co/Kijai...

Neuromasters•9/8/24, 9:13 AM

I tested it, 0 difference and moved to fp8

Neuromasters•9/8/24, 9:13 AM

Errors fail due to memory on the regular version of Dev

Neuromasters•9/8/24, 9:14 AM

I'm just training the lora on he

NNeuromasters I tested it, 0 difference and moved to fp8

magical•9/8/24, 9:14 AM

oh you mean for lora training?

Mmagical oh you mean for lora training?

Neuromasters•9/8/24, 9:15 AM

Yes, on instructions @Dr. Furkan Gözükara

Dreamer@TDFilmStudio•9/8/24, 9:27 AM

I have trained 3 custom Flux loras till now and everytime, I am getting scan lines in images. I use Rank 3 config (slow) with Kohya. Anyone else facing this issue? Not sure what I am doing wrong. I updated to latest Kohya version yesterday, but problem persists. Keeping Lora weight 0.8-1

DDreamer@TDFilmStudio I have trained 3 custom Flux loras till now and everytime, I am getting scan lin...

druhlOP•9/8/24, 9:29 AM

You mean those vertical lines? Can it be an issue with something you're using to process the initial dataset?

druhlOP•9/8/24, 9:29 AM

Even the bottom edge of the generation doesn't look clean, which makes me wonder.

Dreamer@TDFilmStudio•9/8/24, 9:30 AM

I have used 2 different datasets for custom Lora training till now. Each of 20 images

Dreamer@TDFilmStudio•9/8/24, 9:30 AM

Both are producing such lines

Dreamer@TDFilmStudio•9/8/24, 9:31 AM

Even when I use different versions of same Lora saved at various epochs, getting scan lines
Not sure if this is because of overtraining. All images are 1024X1024

Dreamer@TDFilmStudio•9/8/24, 9:33 AM

here another example. Faint vertical lines are visible if you zoom in

Maxivi•9/8/24, 9:38 AM

@Dr. Furkan Gözükara https://huggingface.co/OpenGVLab/InternVL2-8B/tree/main

this is the best captioner i found for flux, tested with 20 images, got my best results with them

OpenGVLab/InternVL2-8B at main

Maxivi•9/8/24, 9:45 AM

26B also nice, think could work on 4bit

MMaxivi @Dr. Furkan Gözükara https://huggingface.co/OpenGVLab/InternVL2-8B/tree/main th...

Dreamer@TDFilmStudio•9/8/24, 9:54 AM

Thhis is the top ranked in VLM models benchmarks

DDreamer@TDFilmStudio Thhis is the top ranked in VLM models benchmarks

Maxivi•9/8/24, 9:55 AM

Yep even the 8b model tops many of the other ones

MMaxivi @Dr. Furkan Gözükara https://huggingface.co/OpenGVLab/InternVL2-8B/tree/main th...

Dreamer@TDFilmStudio•9/8/24, 9:55 AM

Did you use any app or tool to load this model?

Doc @Dr. Furkan Gözükara , I am using default (16 bit) setting for DType on massed compute. When I c

Similar Threads

Similar Threads

Similar Threads