This Lora for SDXL generates digital painting style images of scenes inspired by the medieval period. I tried to get a wide enough range of training images that includes typical medieval scenes like castles, soldiers, marketplaces, and countryside landscapes across different locations. There's a nice range of lighting, locations, and times of day from summer fields in England, to burning buildings in Norway, to opulent Indian palaces.<h3 id="usage-nabcwl7ii">Usage</h3><ul><li>Tested with Forge and A1111. When you add the Lora, use the trigger "middleagesstyle-20240816-1" to activate.</li><li>I primarily use at a Lora strength of 0.8 to 1.0</li><li>Usually, I use the following resolutions with SDXL: 1024 x 1024, 896 x 1152, 832 x 1216, 768 x 1344, 640 x 1536. If I want a higher resolution than that, I'll enable the high-res fix. </li><li>There are different locations (mainly "England", "India", and "Middle East") in the training set, and different scenes with people, but the images are mostly landscapes.</li><li>As it's based on the raw SDXL, it struggles with nudity, but there are obviously many other Loras to help with that.</li></ul><h3 id="training-info-oi19x7tup">Training Info</h3>I started with a lot more, but narrowed it down to 39 images trained to 16 epochs. The training set is concept art digital paintings, partly from a popular medieval strategy game, plus some generations from previous test loras, and other sources.I'm not 100% happy with the exact range of styles. It is mostly is where I want it to be, but occasionally swings too realistic, or too abstract through too few brushstrokes. Portraits of people in general are too realistic, I think. The results are very good, but could be better, and I hope to revisit this, perhaps with the new Flux model.

Medieval Memories

Trained to 39 images over 16 epochs.

Files with SHA256: 8a6a7b47a460e220aa6685cd2141f72de5d7b363a4d869961e79ecf8357d3465