Niji Noob

NoobAI XL v-pred model (requires different settings than usual to get the most of it, see below)

Merged by dumping web UI model from memory to disk

My contribution: I've balanced the painterly look with coherence and used some of my own LoRA's for adjustment.

Settings:

Steps: 14, Sampler: DPM++ SDE CFG++, Schedule type: SGM Uniform, CFG scale: 1.7, Size: 832x1216

Tips:

This is a v-pred model so CFG needs to be very low depending on the sampler/scheduler or you will fry your gens.
Lower CFG (e.g. 1.1) will produce less saturated results. I like my gens saturated so I use 1.7.
You really only need 15 steps at most, less for hiresfix.
reForge has the best memory management of all the SD webUI's so I can output 16?? x 2432 no worries on a GTX 3090 (using 2x on hiresfix).
Use NoobAI artist tags if you don't like / want to adjust the default look.
Non v-pred (as in normal) LoRAs works fine (tbh even better in most cases).
If you get tired of circles in backgrounds the solution is simple: put "tunnel" in your neg prompt.
For more Niji-like output use simple prompts and let the model steer your gens
Want to merge your own similar model, check out the LoRA's used by exlight: AIBooru
Also check out LyiaEngine for CivitAI backups of the above: LyliaEngine (Narontaka)

Description