250D and 50D in this one. 200T and 500T in the other. From my research it seemed that they should look similar apart from noise/grain. As I was fidgeting with my dataset earlier, I realized I should probably try out another strategy. It's impossible to generate a consistent style from an assortment of different photographers, as the processing after capture is at least as important as the film stock. I will instead look for individual photographers with a wide range of subjects and a consistent style, and then make a few lora varieties.

There are no trigger words. From what I've read, that's ideal for a film LoRA as it has lesser side effects. Use it with a weight of 1.0 to achieve the look, but it's possible to raise and lower. I've used a wide variety of prompts and base models to generate the images and tweak the versatility.

I'll update the description of the model so that it's clearer!

That's correct! The reason why the later ones are so small, is that I'm only training the last U-Net block to minimize the effect on composition. As a result, the size of the LoRA is 17 times smaller than it would've been if I trained every U-Net block with network dimension 8.