This lora was trained for 90 epochs using onetrainer on a dataset of 63 high-res frames from Twin Peaks, Mulholland Drive, Blue Velvet etc. It focused on color grading, so no trigger words are required. For caption-based prompting I use qwen captioner to generate prompts for examples.
Works normally on 1.0-1.2, you can mix it with another realism loras as well.