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
seems like discussion here is usually around only 10-20 training images, and that by 1000-2000 steps the lora produces a great likeness (assuming you're going for a character lora). since you have 3x-6x as many training images, you'll have 3x-6x as many steps... steps is images * repeats * epochs. the doc's configs always have repeats at 1, so it's just images * epochs. might be worth additional testing to loras made by 10 images at 200 epochs (2000 steps) vs 20 images at 100 epochs (also 2000 steps). if we increase images, do we get a "good" lora faster? or if we increase images, do we need more steps to better "merge" all of them? i am just a hobbyist and now i'm rambling lol - would be interested in the Dr's feedback.
good news about your electricity - in that case i have more questions lol in your SDXL fine tunes, you train off the RealVis model instead of the SDXL base there are several fine tuned flux models on civitai with thousands of downloads already... should we be able to use those as a base like you did with SDXL? should flux loras work across models?
@Dr. Furkan Gözükara are your sample prompts on patreon anywhere? do you use a wild card extension? would love to try my loras against some of your same prompts.
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
F
F
N
G
