A LoRa model of the BMW M5 (2024), trained on a CGI-generated dataset. The LoRa is highly adaptable, handling a range of perspectives and reproducing fine details effectively. It includes several concepts optimized specifically for automotive photography, focusing on achieving a realistic look and feel. The model performs well from any angle, with minimal background bleed.

It’s designed to work best with a custom ControlNet for enhanced detail reproduction. However, even without this ControlNet, you can fine-tune the details by adjusting blocks dbl_15, dbl_17, and sql_05-sql_13 during inference—ideal values typically range between 0.75 and 1.25. The ControlNet still needs a few more weeks of training and is a bit more complex to use than standard ControlNets, as it requires specifically tuned CGI matcap renders as input. I’m undecided on releasing it yet… we’ll see. That said all images in the post are generated with just adjusting the blocks so it's basically what you can achieve without too much hassle BUT you will see geometry drift in the ranges of 10%. To achieve 100% reproduction you need to work with a ControlNet.

The LoRa includes over 30 integrated trigger words. Some triggers specify perspectives, like “Three Quarter Front,” “Three Quarter Back,” and “Side”; others focus on lighting conditions, such as “Lights On”, "Parking Lights On". Material options include “carbon fiber accents,” “silver accents,” and “black accents.” Additional triggers allow customization of car paint finishes, like “metallic” and “matte,” and even the number plate text, like “number plate reads XYZ.” There are also options for specifying the driver "driver 20 year old wearing xyz" and more — see if you can find them all! :)