Actually, I used 341 images with captions, but the training was only 1600 steps at 0.0001 LR. Maybe it’s too few steps for such a large dataset, but … it’s working, you know?
Work for poses and NSFW, so for me it’s fine :)
(Pero es una salvajada en términos técnicos xDD)
That's so cool - dang, I got a lot of work to do....I'm using Koyha for the LoRA and Stable Diffusion for image generation (I'm trying to create realistic image generator). I have a new found respect for these generators/ai training after trying this - we both know how much work goes into this...How did you caption and verify all 300+ images quickly? Appreciate the tips and great work.
Actually, I used the built-in WD14 Tagger utility inside kohya_ss. It’s much easier than writing a txt manually for every image, especially for 300+ images, because it generates all the captions automatically. I just added the names ‘Anna’ and ‘Starmaker’ to everything. It’s very basic, but it works, as you can see. 😅
