Hello everyone. I am Dr. Furkan Gözükara. PhD Computer Engineer. SECourses is a dedicated YouTube channel for the following topics : Tech, AI, News, Science, Robotics, Singularity, ComfyUI, SwarmUI, ML, Artificial Intelligence, Humanoid Robots, Wan 2.2, FLUX, Krea, Qwen Image, VLMs, Stable Diffusion
Hello, I’m still a beginner and learning through SECourses, so I apologize if my question is unclear.
I successfully trained a character using Flux Full Fine Tuning (or DreamBooth), and I tested the result in ComfyUI as a UNet model (without using flux1.Dev). The results were very realistic and successful.
Now I’m wondering:
Can this trained model be used as a replacement for Flux Kontent?
Or, can I use Flux Kontent and combine it with my trained model as a LoRA?
I want to fine tune on a specific quality type of images of women. Not 1 character, but I want the fine tune to be able to generate woman with same quality as training data. Say I have pics of 5 to 10 different women of multiple shapes, hair, body, ethnicity etc. How many pics would be needed for a high quality finetune? Ideally it should be able to generate women not in the training set, I.e. an Arab woman, but with same quality as training pics. Unsure how to proceed, only done fine tune with 1 character so far. I guess captions should be needed here.
Hello! I am trying to get the kaggle notebook to run "kohya-sdxl-lora-training-on-a-free-kaggle-notebook_v28", but the moment I import it just starts downloading a ton of dependencies automatically. I did not see this in the tutorial. What is the expected behavior? Thanks in advance
Hey, just wanted to throw this here FYI in case you're looking for STAR alternatives within your APP, you can check if you want, they claim to be better than STAR, however, I haven't been able to test it out. https://github.com/yjsunnn/DLoRAL?tab=readme-ov-file
I'm using the configuration you have been using for 48 gb ram. I rented an H100 GPU with 20 image dataset but the images generated after training does not even match with my training data. Where do you think I might be going wrong? all the parameters are loaded from 15 image dataset config file. Can you please advise?
yes I'm loading them in the dream booth tab. I also am loading the training images properly but just to be sure is there a way for me to know if the training images are being loaded properly or not?
/kaggle/working/kohya_ss/training_data/img/1_ohwx man... 00:37:48-667704 INFO Regularization images directory is missing... not copying regularisation images... 00:37:48-670635 INFO Done creating kohya_ss training folder structure at
and when I start training, you can see them in the message.txt file.
I tried other images including mine but didn't work, so I tried taking a few screenshots from the 15 image dataset of yours and at the 1000th step the sample image I get is not even close to the training images. I will try with Dwayne Johnson images now and give you an update.
I'm not sure either, because the interface I'm using is jupyter lab but the folders being created says kaggle. Let me try using massed compute and I will update you with the results I get. Thanks for your advise.