Software Engineering Courses (SECourses)•2y ago

Yep i trained my own base model, I probably will release it soon when its done, there are still some

Yep i trained my own base model, I probably will release it soon when its done, there are still some tweaking needed. I still have another 5-6k images left to train. I captioned those images with llava.

then i train my own models, celebs etc on it. Comes out quite nice

DJRich•12/31/23, 5:54 PM

Has anyone seen this before? I have a laptop with a 4090 and 16GB of VRAM. Out of the blue I am starting to get out of memory errors during a hires fix in A1111 where the upscale is only 1300x1700 approximately. The same images worked a couple of days ago without issue. Not running any updated versions of any software and even switched from sdp to xformers because that seemed to help upscale some images that were failing. According to Windows 11 my GPU is using 8.4GB of VRAM just starting A1111. I've tried disabling all extensions and no change. Without A1111 running VRAM is 0.3/16GB.

FFiveSAI Yep i trained my own base model, I probably will release it soon when its done, ...

Lorelai AI•12/31/23, 6:07 PM

do you have any good resources on llava I have been looking at decent captioning. Blip I find not that great nor is git tremendously accurate

DDJRich Has anyone seen this before? I have a laptop with a 4090 and 16GB of VRAM. Out o...

Furkan Gözükara SECourses•12/31/23, 6:10 PM

interesting

LLorelai AI do you have any good resources on llava I have been looking at decent captioning...

Furkan Gözükara SECourses•12/31/23, 6:10 PM

i have auto llava installer

Furkan Gözükara SECourses•12/31/23, 6:10 PM

batch caption as well

Furkan Gözükara SECourses•12/31/23, 6:10 PM

https://www.patreon.com/posts/sota-very-best-90744385

Patreon

SOTA Image Captioning Scripts For Stable Diffusion: LLaVA, BLIP-2, ...

Patreon is empowering a new generation of creators.
Support and engage with artists and creators as they live out their passions!

LLorelai AI do you have any good resources on llava I have been looking at decent captioning...

FiveSAIOP•12/31/23, 6:10 PM

I used the same one he posted on his Patreon for the auto installer. I had used a previous one before.

I only use llava for fine tuning base models.

I like blip for training people. With an instance prompt

FFurkan Gözükara SECourses i have auto llava installer

Lorelai AI•12/31/23, 6:10 PM

Awesome

FiveSAIOP•12/31/23, 6:11 PM

If you caption wrong you’ll get bad results

FFiveSAI I used the same one he posted on his Patreon for the auto installer. I had used ...

Lorelai AI•12/31/23, 6:11 PM

I don't understand blip entirely but I found that the higher the beams the worse it got

FiveSAIOP•12/31/23, 6:12 PM

I usually edit them. Yeah I stopped using higher beams, I just use 1 now. And I change the other setting from 0.9 to 0.7

FFurkan Gözükara SECourses interesting

DJRich•12/31/23, 6:12 PM

I just installed the A1111 package from StabilityMatrix with all of the same extensions and it only takes up 4GB of VRAM.

FiveSAIOP•12/31/23, 6:13 PM

Things I’ll usually caption is just my instance prompt, their hair style, and length, and background.

FFiveSAI Things I’ll usually caption is just my instance prompt, their hair style, and le...

Lorelai AI•12/31/23, 6:14 PM

I also caption anything I wouldn't want in the shot because I feel if the training is inclined to add it I could aways put it in the negative prompt

FiveSAIOP•12/31/23, 6:17 PM

Also I have started testing removing the backgrounds from my images, also decent results

FiveSAIOP•12/31/23, 6:17 PM

Idk I’ve never got good results using no captions

DDJRich I just installed the A1111 package from StabilityMatrix with all of the same ext...

Furkan Gözükara SECourses•12/31/23, 6:18 PM

you can also add --medvram and --lowvram

Furkan Gözükara SECourses•12/31/23, 6:18 PM

to auto1111

Furkan Gözükara SECourses•12/31/23, 6:18 PM

will have same effect

FFiveSAI Also I have started testing removing the backgrounds from my images, also decent...

Lorelai AI•12/31/23, 6:19 PM

I need some insight, I am not normally going to train outside of people or anime characters. I want to do Namek before it's getting ready to explode from Dragonball. My questions would be, how many pictures of the planet Namek in that state, should I remove any characters with a magic brush, and Ideally I imagine the class is planet, what type of planet regularization images would I use and where can I get them?

LLorelai AI I need some insight, I am not normally going to train outside of people or anime...

Furkan Gözükara SECourses•12/31/23, 6:20 PM

just skip reg images

Furkan Gözükara SECourses•12/31/23, 6:20 PM

in harder to get cases

FFurkan Gözükara SECourses to auto1111

DJRich•12/31/23, 6:20 PM

Yes, but I really shouldn't be seeing this issue with a 16GB card not for such a small hires fix. Especially since it worked perfectly before. I will try the StabilityMatrix version for a while and see if it is more stable than before.

FFurkan Gözükara SECourses just skip reg images

Lorelai AI•12/31/23, 6:20 PM

that's a good idea

LLorelai AI I need some insight, I am not normally going to train outside of people or anime...

FiveSAIOP•12/31/23, 6:20 PM

Outside of people I don’t use reg images

Lorelai AI•12/31/23, 6:21 PM

What about people in the shot if I can't get enough training images without. And how many training images? 14 just like people?

FiveSAIOP•12/31/23, 6:21 PM

Well for backgrounds you could get away with just using 10 images

FFiveSAI Well for backgrounds you could get away with just using 10 images

Lorelai AI•12/31/23, 6:24 PM

Awesome! There's actually a bunch of good images online for Namek's destruction I should be good there

LLorelai AI Awesome! There's actually a bunch of good images online for Namek's destruction ...

FiveSAIOP•12/31/23, 6:24 PM

You can probably rip some from the old dragon ball games

FFiveSAI You can probably rip some from the old dragon ball games

Lorelai AI•12/31/23, 6:25 PM

I'd get them from the actual show if i had to

FiveSAIOP•12/31/23, 6:25 PM

For sure, you could probably dabble in img to img, to upscale and get better quality

FFiveSAI For sure, you could probably dabble in img to img, to upscale and get better qua...

Lorelai AI•12/31/23, 6:27 PM

I use pixelcut, you actually helped me refine most of my process. The biggest tweaks I changed though were I don't use blip and pretty much manually caption the git captions and I use a training rate of .001, that was a huge improvement. And I thought .001 was high, but documentation says that AdamW is optimal at .001

Lorelai AI•12/31/23, 6:29 PM

learning rate I mean

Sathish H•12/31/23, 6:36 PM

Hi All, One question, i want to train stable diffusion to create pictures with background of local tourist places. Any suggestion on how to train, technique etc.

SSathish H Hi All, One question, i want to train stable diffusion to create pictures with b...

Furkan Gözükara SECourses•12/31/23, 6:44 PM

you need to collect lots of good images

Furkan Gözükara SECourses•12/31/23, 6:44 PM

then you can llava caption them and train

Furkan Gözükara SECourses•12/31/23, 6:45 PM

or with activation token + caption

Furkan Gözükara SECourses•12/31/23, 6:45 PM

you should experiment

SSathish H Hi All, One question, i want to train stable diffusion to create pictures with b...

Lorelai AI•12/31/23, 6:56 PM

I'm about to do my first one I can give you feedback on how it went

FFurkan Gözükara SECourses or with activation token + caption

Sathish H•12/31/23, 7:26 PM

Thank you

tom•12/31/23, 9:28 PM

Hmm. Tried a few times. I keep getting the following error : orch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 MiB (GPU 0; 14.76 GiB total capacity; 12.34 GiB already allocated; 3.75 MiB free; 13.47 GiB reserved in total by PyTorch. I've gone down to train just 9 1024x1024 images with this same error. Is there a way to flush the gpu or something? cudnn doesn't register.

tom•12/31/23, 9:28 PM

I'm re trying with lower repeats. Fingers crossed

Kallamamran•12/31/23, 9:40 PM

Happy New Year y'all!

Ttom Hmm. Tried a few times. I keep getting the following error : orch.cuda.OutOfMemo...

Furkan Gözükara SECourses•12/31/23, 9:59 PM

where are you trying what are you trying?

FFurkan Gözükara SECourses where are you trying what are you trying?

tom•12/31/23, 10:23 PM

Following the newer video. Instead of the male cell, used the female cell. Added 100 input images. Failed. Then tried again with 9 images. Failed. Then restarted everything with 9 images and no Regularisation images. This too failed Cuda out of memory error. Tried it again and it worked in about 36 min. (Obviously, this is just a test and the result not usable) I'm wondering if Kaggle is having some issues. Had errors yesterday as well.

tom•12/31/23, 10:23 PM

I'm going to try again shortly with a proper training. Fingers crossed.

tom•12/31/23, 10:25 PM

Any plans for a Comfyui via Kaggle class?

Ttom Any plans for a Comfyui via Kaggle class?

Furkan Gözükara SECourses•12/31/23, 11:18 PM

i tried but it didnt work on kaggle

tom•12/31/23, 11:46 PM

Ok, I will make an attempt at some comfyui via kaggle code after this training is done and let you know.

mjak•12/31/23, 11:47 PM

How does everyone here handle the repeats/epochs when doing the SDXL Training with Textencoder training?

The formula that @Dr. Furkan Gözükara mentions is:

Max Train Steps = Train_imags * 2(if class img used) * repeating_count * number_of_epochs

In the medium article @Dr. Furkan Gözükara mentions that it's 4500 steps at 150 repeating but using the formula above with regularization images = 9600 steps (15x2x40x8)

My dataset is 20 images so I've done 35 repeats per image and at 8 Epochs = 11,200 Steps
(20 images x 2 x 35 x 8)

Yep i trained my own base model, I probably will release it soon when its done, there are still some

Similar Threads

Similar Threads

Similar Threads