Get all 100+ of my premium LoRA for free by joining my membership including Karina v12 on my kofi or buymeacoffee. If you want to make your own custom LoRA you can also contact me there.
Trigger word: aespakarina
Best checkpoint to generate the images for V8 and newer versions: ALEREAL
Recommended settings to use for v11 and newer version:
- Sampling method: DPM++ SDE Karras
- Sampling steps: 25 or more
- Use Hires fix: 8x_NMKD-Superscale_150000_G or 4x_foolhardy_Remacri.
Hires steps: 10
Denoising strength: 0,4
CFG scale: 7
- Negative prompts, always use this to get best outputs:
(paintings:2.0), (sketches:2.0), (worst quality:2.0), (low quality:2.0), (normal quality:2.0), (lowres:2.0), text, watermark, signature, artist name, rendering, 3D, painting, semi-realistic, anime, cartoon, unrealistic, fakeUse it wisely. I'm not responsible for anything you generated with this LoRa.
Join KPOP Diffusion Discord: https://discord.gg/NrBDMX79Da
Description
2nd version. Trained on latest version of Linaqruf's colab using 30 images with HassanBlend 1.4 as its base model.
I use new settings for this, it's better in realistic and face similarity than the previous settings but it's harder to stylize it especially hair colors and painting style. Best at close-up, pose will be repetitive at some seeds and prompts because its dataset have many similar pose, but it can be solved with Controlnet, the face similarity will be decrease though.