From the legend himself, Grigbertz! I've been following his work for nearly 2 decades now and figured Id make a LORA. Ive only used colored images for training data :)<hr />I've found a great technique that maintains the flexibility of using score tags while preserving the visual fidelity of the specific artist’s style:<ol><li>Generate with Score Tags: Start by using score tags in combination with the style LoRA at a strength of 1.0. Continue generating until you find a seed that works well.</li><li>Ensure Face Detail Consistency: If you're using a face detailer, make sure it's also using the style LoRA. Otherwise, the face might end up looking airbrushed or awkward, which can detract from the overall style.</li><li>Upscaling with Style LoRA: When sending the image to your upscaler, use the style LoRA again, but at a slightly higher strength (around 1.2) with a denoise level of 0.5. Change the prompt so that it removes all score tags. This helps to overwrite any unwanted stylistic effects from the score tags and ensures the final image matches the desired artistic style.</li></ol>That’s all you need! Give it a try and let me know how it works for you.

Files with SHA256: 54ce9d17975d24df74503456c05ae83d1c810ee54a288840053a4735ff81a9ad