Thanks to a training workflow kindly provided by 7err4 ,this version actually produces different outputs from the base model, increasing the realism of images across all sampler types and requiring less prompting for a realistic style. However, some character's/concepts may still have a plastic or overexposed look due to some of the images I used in the set being AI generated or CGI. I recommend using negatives and additional prompting to try and rectify this issue, but I will have to prune my dataset in hopes of making future versions work better. The trigger word is just "ph0t0an1ma" and it works fine enough to trigger the style, though I recommend using "cosplayer" to describe characters you want as that seems to nudge the base checkpoint toward realism. Since the base model is capable of some realism you can probably lower this LORA's strength if you think its altering the proportions or textures too much.
