yet another question about training
Hi, first of all - great work!
May I have few questions to you?
As I have some spare GPUs - which one would be beneficial for you to train on? What do you actually train this on? Have few 5070 Ti, 4090s, 3090s... probably not what u are looking for but in like 4x setup they are kinda capable..
Do you plan to release your training script like how you achieved you could use this input - control + caption dataset and that lore actually reacts to it? And I am really impressed you trained this on such a small dataset? Like upto 2k images? Thas crazy..
How many epochs? 2-5? Crazy..
Would much bigger dataset actaully help you? Like tens of thousands of captioned pairs or that only hurts the lora? But how it can learn to generalize that well if it might not see lot of things at all? :)
You can try this code I made, or you can try creating a patch for the ai-toolkit from it if you prefer. There's a document that explains how this training works.
Training code is now public: https://github.com/lbouaraba/krea2edit-trainer — ai-toolkit extension, one git clone into extensions/.
Measured VRAM peaks are in the README: r64 @512 = 28.1 GB (comfortable on 32 GB cards), 24 GB works with cached text embeddings; 16 GB is below the current floor, with a low-VRAM mode tracked in the repo issues. Your 3090s/4090s can train this today. :)