The model can reproduce the BYD Yangwang U9.
This is my most accurate model so far. The training workflow is a bit more complex; it was trained with 3 datasets in different resolutions and various repeats using Kohya.
For the captioning, I used CaptainCaption from 42lux, which creates very precise captions with the help of ChatGPT.
Many thanks for the tool and the tips to 42Lux!
The Yangwang U9 is a battery electric sports car manufactured by BYD Auto. The U9 is the second vehicle under the subsidiary luxury car brand Yangwang, and was formally introduced on Auto Shanghai in April 2023
Description
FAQ
Comments (4)
Hi, really a great result. I am also very interested in the training strategy with multiple datasets at different resolutions and repeats that you have adopted.. would you like to share the details?
The Idea is from 42lux. U use the same Images in 3 different resolutions (1k,2k,4k) and train the lora with different repeats, low res with the most repeats and high res with less repeats.
@denrakeiw thanks for the super quick reply! but I was wondering: doesn't the bucketing system have a hand in it? Meaning: doesn't it scale 2k and 4k images back to 1k before training?
@gattaplayer Yes, there is an option "Don't upscale bucket resoluion" u have to check this.