i was asking deepseek about it, and seems that you're right, for having the best captioning you need to do a forensic analysis of the pictures, even more on videos, ill try to do it better on the next ones..😥 atm im training one with just keywords, not even single words, keywords, like deepthroat_fast_ontop, but i guess it will have the same issue..if only we could captioning somehow automatically with hight level of detailed descriptions.. xD

Btw im open to do a collaborative lora, if you prepare the datasets and the captioning i will use my hardware to train It..

@Degenerator123 yea i saw those and i tested but i was getting errors and wasnt able to run it, so im using Shutter Encoder for processing videos, and an excel with macro for exporting the captions in one batch y just need to caption every aspect of video per column and every clip per row in the excel and then run the macro, works quite well but you still need to writte and manually describe, i will try some python script DeepSeek gave me

yep, is literally the double or even more the amount of time but i guess is worth it for the results looks better, im still learning tho

@Baka_Oppai i meant, the closes to 1 learning rate, the loss to steps relation is slower, so for the lora to get a 0.08 - 0.05 loss (thats when its ready, right?) will take more time with 2e-5 learning rate, i mean i realized that when i went from 5e-5 to 2e-5 and observed the TensorBoard graphs.

@Baka_Oppai as far as i know, the diffusion-pipe that i downloaded (the WSL ubuntu image resource in civitai), doesnt allow you to modify the learning rate parameters once the train is launched, gpt and deepseek told me that much.. i read something about a dynamic learning rato but i needed to modify the base script "train.py" so i pass.. im seeing ppl doing great loras with 2 learning rate... so i guess i'll stick to 2e-5, the new lora i released i just uploaded it at 40 epocs 9700 steps, im still training it, the graph looks still recuding the loss, so i guess i will again post the same lora but with more steps/epochs like on my first lora,, i will share the graphs when i upload the v1.2

@logenninefingers888 awesome info, im at 70 epochs right now for v1.2 of my new lora, i Will share the graphs of It doesnt overcooks😅, this is actually the longest train run i did

thanks for your kind words :D

I mean all feedback is welcome, but some here tend to act like they are renting me their computer and the electricity bill is on them...I mean...I wanted to get the deepthroat part, not the nitpicky little detail, theres a lot of blowjob loras here but i allways wanted them to go deep, ... so i created a deepthroat one, not a character creator.

you can use those as well, i duplicated, swaped from left to right the image, and recaptioned, then used both datasets.. so you can use those as well