i also include the color, sometimes the captioning does that for me, i also add "front view", "viewed from behind" or "side view" to further make the ai understand what to create, but sometimes it doesn't work but it's worth a try, i think larger data sets are required

i use 15 high resolution images and also upscaled on gigapixel ai, sometimes the upscaling isn't even required if the base images look really good.

i search for the images on google and look for them to be at a high resolution, this time i also searched on instagram because i already knew i would find some very good images of a rx7 compared to google where i'd have to search a bit harder.

i also enable buckets because i dont want to resize my photos and lower their resolution and quality.

also because i try to get high resolution and good quality pictures sometimes the background is the same because the photos were taken in the same photoshoot and keeping the background the same will teach the ai your background to be associated with the lora and that kinda sucks to be honest... i have found 2 solutions to that: photshop and either copy the car to a new background or copy a new background onto your car (requires a lot of work and is time consuming); lower the strenght of the lora but that also might reduce the fidelity and realism which i try to keep...

i now train 3 epochs each has 1500 steps and try to see which one learned the best, takes me about 40-60 minutes to train plus the image gathering +/- upscaling

@MyNameIsThis thanks for the tip