Remember to use the trigger word parksooyoung. Use LORA weight 1 and Hires fix with latent bicubic antialiased, 20 steps, 0.5 denoising strength, and 1.25x upscale. (I've also been liking high cfg, e.g. cfg 15, with dynamic thresholding, though this definitely needs face inpainting.) I've used <a target="_blank" rel="ugc" href="https://civarchive.com/models/6424/chilloutmix">ChilloutMix</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/4834/visiongen-realism-reborn">VisionGen</a> to generate images.

Trained completely differently compared to my last two LORAs. Different training settings, much bigger dataset, and a little sauce. All this work, yet I'm not sure it's much more flexible or accurate than how I previously did it, for example, it really likes the side view/looking back despite cutting down on those pics in the dataset. As per usual, inpainting and ControlNet should help a lot.

Files with SHA256: fa81872039b64b02ff262d6af09d07eb37161f4dea2be1ba9e9f38c638af998d