Hi everyone,This is an experimental LoRA I created, and I wanted to share both the LoRA and the training data for download. Here's a quick overview of what I did:<h3 id="experiment-process:-i6qxq0usz">Experiment Process:</h3><ul><li>I gathered 100 images from MidJourney's Top of August collection. These images were diverse, including various sizes, aspect ratios, and styles.</li><li>I used the MJStyle824 trigger word.</li><li>The images were captioned using Florence 2 Large in Detail_Description mode .</li></ul><h3 id="-vk8hx9ci9"></h3>This was purely a curiosity-driven experiment to see if any recognizable style or effects would be captured from such a small and inconsistent dataset. The results were surprisingly noticeable, given the dataset's limitations.<h3 id="effects-and-performance:-si09vmjjs">Effects &amp; Performance:</h3><ul><li>The effects this LoRA adds are relatively minor, but still more than expected considering the limited and inconsistent data.</li><li>The LoRA works best with higher strength settings in the 0.8 - 1.5 range.</li><li>You can safely push the strength up to 2 for more abstract or non-realistic prompts without encountering significant issues.</li></ul>Enjoy this quick little experiment and please share, if you get anything interesting :)

Files with SHA256: e0d5c0b185bf87a84caa052e5f06ad0e3b241cc82dec866011f896887cd52e6a