If you want to learn how to train faces with high fidelity, reach out to me on my Patreon. :)

About:

The purpose of this LoRA is to (try to) closely replicate the appearance of Shiftymine, a famous Twitch streamer.

It was trained in multires (512 up to 1024), using simple captions and with 46 high-resolution images.

The demo images of the LoRA aren't edited, but they are cherry-picked and generated using hires-fix (0.25 ~ 0.45 denoising).

Suggested Settings for Inference:

Model: epicrealism_pureEvolutionV5 (but works with 97,2% of models)

Positive: a woman, full scene shot, winter clothes, hyper realistic absurd, 1 0 / 1 0, heavily upvoted <lora:Shiftymine_v1_SD15_by_IsnAI:1>

Negative: (nsfw, ????, topless)

Sampler: DPM++ 2M Karras

Steps: 32

CFG: 5 (5 ~ 8)

Resolution: 512x768

Upscaler: 4x_NMKD-Siax_200k

Note¹: Applying high denoising values (0.45~) in hires-fix enhances the appearance fidelity.

Note²: I have no clue why, but the tags streaming on twitch and twitch streamer produce pictures of people with their tongue out and/or mouth open.

Triggers:

Main: a woman

Angle: full face shot, head and chest only, full scene shot