Ver2ish:
Attempted to improve prompt following and get rid of some of the ocasional multi-limbed eldritch horrors. Sample images generated using ComfyUI single pass without high-res fix (mostly using either ChatGPT or a local LLM to improve the prompts)
It's capable of some NSFW, but has the usual issues with ... lets call it intimate anatomy.
Suggested Generation Parameters:
Sampler: Euler_a
Scheduler: normal
Steps: 8
CFG: 1.2
________
Lightning8 Beta1 - 23 Feb 2024:
A merge of FastPass core and Lightning8 step.
Suggested Generation Parameters:
Sampler: Euler_a
Scheduler: normal
Steps: 8
CFG: 1.2
____
This merge combines a handful of my favorite realistic models and Lora, along with both LCM and SDXL Turbo Lora extractions. It's fast. Blazingly fast. It's also sometimes a little fickle, and ... set in its ways.
Performs best at very high native resolutions - 1024x1360 using Kohya's Deep Shrink Hi-res fix and very simple negative prompts.
Best Settings:
Steps: 4-8
CFG: 1.3-1.6
Resolution: 1024x1360 or derivatives
Generation time 2-5 seconds on a 12GB RTX 3060
I haven't tried many Loras with it, so YMMV.
Description
Updated with new model versions, additional LORA and reweighted for better photo quality, better prompt adherence, and (hopefully) fewer extralimbedeldrichhorrors.
Generation Parameters:
Sampler: eauler_a, eauler_ancestral
Scheduler: normal
Steps: 8
CFG: 1
Works best with short concise prompts, and brief or empty negative prompts. Because of the super low CFG, negative prompts can bleed through into the image, so be sparing/experiment.