[LuisaP💖] TRAIN Z-IMAGE SLIDERS FASTER THAN LIGHT [ 5 MIN ] - OSTRIS AI

So, i will explain why i'm flooding the civit ai with a lot of sliders loras.
it's because no data is needed at all, i will explain soon.
first of all, you need to modify the ostris ai toolkit to be able to train.

go to /toolkit/prompt_utils.py
change text_embeds = list, to text_embeds = padded
if you are using runpod like me, open web terminal and place this.


sed -i 's/text_embeds = embed_list/text_embeds = padded/' /app/ai-toolkit/toolkit/prompt_utils.py

make a dataset with random images, can be a bunch of people, don't matter, no caption needed.

disable TE, set rank to 4, lr 0,001, steps to 200 and save every 50, and the rest become normal, no need to change, maybe the image sample you need to change the prompt, and make a -2 and 2 lora strenght to see if is working.

BUT

in resolution, change this

to this

this do the deal, almost 5 minutes on low vram.

good training chooms!