Recommended parameters:
Sampling method: DPM++ 3M SDE (depending on the model, otherwise Euler A)
Schedule type: Exponential (depending on the model, otherwise Karras)
Steps: 40 (depending on the model, otherwise 30)
Upscaling: None (Native, depends on the model. Go 1800p or 2048p)

Why you should avoid HR Fix and Upscalers
Upscalers are all terrible. They smear and destroy fine details, turning them into goo the higher you go. If you absolutely need to upscale, do it after generating at the highest native resolution possible. You will also save more time generating your images and get to upscale only those you like.

Only DPM++ 3M SDE and Euler A can support resolutions up to 2048p without introducing early distortions. However, be warned that Euler A is a stochastic ancestral sampler, meaning every image generated with it is inherently random and cannot be perfectly re-created, even with the same seed and prompt.

Not every model can handle 2048p. If you notice malformations, stretching, or distortions, try Kohya DeepShrink or Kohya HRFix. If those still do not fix it, your model simply does not support that resolution and you will need to lower the native resolution or switch to another model.

I hope this helps anyone trying to get the best possible results! Not just for this LoRA but for all your future images as well!

About this LoRA:

Sorry, I took so long. I meant to upload this LoRA three weeks ago. My first LoRA project! Trained from scratch using only 5K UHD trailer frames. I redrew and repainted nearly every shot to restore missing lines and colors, aiming for complete consistency despite the limited data. I learned that tagging every defining trait, pose, and action is crucial for stability. Lykaia drives most of the time in one trailer, so her gestures had to be treated as learned actions rather than defaults. Otherwise, her arms would be up as if she were holding an invisible steering wheel, and she would be in a driving/sitting position, as if she were driving an invisible car in every generated image. I know this because it happened. lol

Beginner’s luck was definitely on my side; her core appearance was nearly perfect. Unfortunately, I accidentally overwrote the original LoRA while trying to fix her outfit, which forced me to start over. From there, I found myself caught between conflicting advice and guides from people who, I believe, do not fully understand how LoRA training actually works. (No offense to those who do).

If the hair ornament or hairclip bothers you, try adding (hairclip:2.0, hair ornament:2.0) at the beginning of the negative prompts. I will be removing it in the next training pass so people can freely customize her look with their own accessories. I have already tried training closer shots of the Nova hairclip for two weeks straight, but it makes no difference. In fact, it worsens the overall proportions by enlarging the head and distorting perspective. The only real fix is to gather many more images that clearly show the hair clip from consistent angles and without distortion. And for that, the character itself needs to be closer, not cropped or zoomed in. Cropping or zooming 5K frames only made the linework thicker and blurrier, which the AI learned was not a good mix.

As for people asking why I did not use Danbooru, DeviantArt, or Pixiv, etc, well, try searching for her yourself, you will see. There are absolutely zero fanarts of her anywhere online despite all the love she got at first sight! Shocking, I know, I was too! Either no one feels like drawing her, or no one wants to because there is still no clear reference image of her updated design. Not even 3D models.

Anyway! I will keep refining this LoRA whenever I can, so stay tuned for updates!