I used the RealisticVision V6 B1 as a starting point for this model as I was hoping to achieve a higher degree of photorealism compared with my previous effort with my airtightintrrcl LoRA which was trained on the Clarity V2 checkpoint. This model is the result of finetuning RealisticVision V6 B1 on a curated dataset consisting solely of high resolution photographs hand cropped to a 3:2 aspect ratio and scaled down to a 768x512 resolution. As a result, the model is unlikely to produce impressive results if you attempt to use it to generate images at other aspect ratios / resolutions (at least in terms of generating the intended content on which it was finetuned).

Overall, I'm pleased with the results. The frequency of generated defects is still higher than I would like, but with the right prompt the model can produce good images at a relatively decent rate. As a result of training on a larger dataset alongside a much larger regularization image set custom generated using the the base RealisticVision model specifically for this training, this model is capable of a more variety in the images it will generate and the level of detail and the degree of photorealism that can be achieved has been improved substantially when compared with the results of my previous LoRA effort.