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
@Furkan Gözükara SECourses, any news on Flux Dreambooth finetuning recommended number of steps? The 17500 steps (175 training images) training I ran the other night is disappointing… I trained on one person and the ohwx person in generated images only bears a passing resemblance…
Has anyone had issues with the vae when trying to finetune in kohya? It looks like it's loading the full flux dev checkpoint as the vae rather than the vae I'm pointing it to in the config...and so it fails to generate latents. I think I can work around this by just not caching to disk and by turning off "use latent files".
Just wondering if anyone has seen this. kohya_ss is all up to date.
I thought I was using your settings, but maybe I increased the repeats in addition to the epochs to get more steps? I can't recall 100% but I guess it was a mistake on my part
This is the issue I'm seeing...I'm pointing it to your vae now, but when kohya runs you can see in this output that it keeps trying to use the flux checkpoint as the vae. For some reason it's not use the vae in the gui or in the toml