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
I must thank you. i used your tutorial to finetune SDXL using a very specific style. (400 training images) but between epochs 75 - 150, they all look good
@Furkan Gözükara SECourses I have trained a finetuned SDXL checkpoint to learn a specific style. The results are excellent.
Now, I would like to finetune that same stylized checkpoint to learn a specific character. Then I will extract it to a LORA. The character is a female demon. But the training images (40 of them) are in various styles such as 3D, anime style, cartoon style.
If I use the caption "ohwx 1girl" or "ohwx woman", I am afraid that character will OVERWRITE the style of the finetune that I want to keep. How do I caption the character so that it does not overwrite the style?
I did a character finetune, and noticed that almost all training images had ponytail hairstyle. I manage to produce some images without the ponytail, like prompting "messy hair", and even "no ponytail", but the latter didnt really work, haha
So at the moment in my situation Lora training is about 3.58s/it, but if I use DreamBooth instead TIER 1 it will be about 7.3s/it??? as I can see on above picture. So double the time. It's worth you think?. Regarding quality?.
With 60 training pics for character, is 200 epochs and save every 25 a good setting? I don't think you changed anything about learning rate and other parameters when I watched your tutorial video @Furkan Gözükara SECourses, just use config batch 1? I'm gonna use a80 or h100 later tonight . Want to complete it as fast as possible, with highest possible quality.