This LoRA will render a BMW M5 into a scene. While this was trained against SDXL 1.0, do give it a try with other model checkpoints! It continues to surprise me each time I use it with different models and combinations of LoRAs as well. So experiement with this one. I have noticed that Euler ancestral, lower guidance and steps seems to produce some of the better results. Your mileage will vary.
Definitely use v1.1 (note that I didn't name the file differently, sorry) over v1.0 as it's more accurate. The dimensions and the details are just better. However, you may find the original version to be a little nicer for illustrated or stylized effects. The first iteration was a bit more difficult to work with and I don't think it was as accurate. The difference is mostly in the captioning, I completely redid the captions for v1.1. The training was slightly different as well. You can try the older version if you like, but I think v1.1 is superior. I may put out a version 2 that is trained off even more images. Version 1 is trained based off 36 images. It is a LoHa using the Prodigy optimizer with a network rank DIM and alpha of 32.
Description
Slight tweaks to the training, mostly changes with the captions to use more of a tagging approach than a descriptive sentence based caption. This seems to have greatly improved the details in the front bumper and grill of the car. However, there may be some slight drawbacks for creative freedom and illustration type images might not be as good as a trade off. After training several LoHas and comparing them, this version has the best balance and is most accurate (even though some of the other versions had twice as many steps, this one was the best for accuracy).
Future versions will likely be new major versions and include the use of more training images. For the first version, this is pretty close to final and where I was hoping to get with things.