Maybe not just another checkpoint merge. My goal was to keep details as crispy clean as possible without looking like the typical 2.5d models out there. It focuses on keeping ??????? consistent and reliable while cleaning up details and producing less... nonsense.

VAE is NOT baked into the model. Suggest to use SDXL_VAE.safetensors

Alternatively, try out SDXL Anime VAE decoder Only B1.safetensors

🎛 Recommended Settings
Sampler: DPM++ 2M, DPM++ SDE, DPM++ 2S a, Euler a
Scheduler: Karras, SGM Uniform
CFG: 4-7 (idk, go wild)
Steps: lots, probably
Hires fix: yes, do it.

At the end of the day, it's an Illustrious model. All your characters and artist tags work for influencing the generation just as they would on the base model.

What I like to do? Generate the base image, use face detailer, upscale with hires fix, then run it through img2img with the same prompt at like 0.2-0.3 denoise strength

Lyrixa v1.1

Like v1.0 but better. Lighting, details, saturation, cohesion.

Lyrixa v1.0

We started here. Probably.