sd-dreambooth-library

/

cat-toy

like 3

I have that question too

Hi how are you i trait to train the IA with the google colab book but first i have a problem with VRAM for 200 Mib less give me an error, i solve it when change 512 to 256 the resolution and 400 to 200 the steps. Finally i could make the training but i havent got good result, i think i make a mistake with prompt, someone who have good result from training could help me please

Hi! I suppose that it depends of the GPU that was assigned to you in the colab, maybe work if you terminate the session and start over again, you can check the gpu with the command

!nvidia-smi

and check if the Vram is enought, 15gb of vram should work...

What i haven't figure out is how to transform the output into a single ckpt, I was check the documents on tensorflow and pytorch, but I have no clue how to find it...

Hi! I suppose that it depends of the GPU that was assigned to you in the colab, maybe work if you terminate the session and start over again, you can check the gpu with the command

!nvidia-smi

and check if the Vram is enought, 15gb of vram should work...

What i haven't figure out is how to transform the output into a single ckpt, I was check the documents on tensorflow and pytorch, but I have no clue how to find it...

+1 on ckpt file, please help =)

Thanks for your answer and work MvsSrs, i do it google colab give me 15109 Mib on GPU but get out of memory but i resolve making that changes, i have a question when i want to make a training of myself what kind of photos or quantity is good, now i try with less photos and only 350 steps resolution 512 and is going fine

For a person maybe 10 to 20 photos should do nice results, in the form of
2-3 full body
3-5 upper body
5-12 close-up / face
however the colab says that 3 - 5 photos should work, and 512x512 looks right since the original model was trained with that resolution...

Thank you! How i can use the library that i create before training in other google colab i know that isnt a way to convert on ckpt file yet

there must be a way to transform from diffusers to tensorflow ckpt, but I haven't found it. Here are some links that maybe help for someone with more knowledge:

https://github.com/jsksxs360/bin2ckpt
https://github.com/huggingface/transformers/issues/9204
https://github.com/huggingface/transformers
https://huggingface.co/docs/transformers/converting_tensorflow_models

I'm a little bit lose, so if anyone knows more about the topic, it would be great :)