Since SD3 has been kinda disappointing, here is my latest merge. It is based on Datass Rev 3, mixed with some other Pony Merges and finally mixed with my favorite SDXL model. Spent a lot of time finetuning and merging CLIP as well.

It can do creative work with the proper prompting, as well as realistic photos and NSFW.

I suggest to use the FULL model, it is way better than the hyper version. You can always just use the optimization LoRAs with the full model as well.

ComfyUI Workflow:

-> Here (18+ GB VRAM)

Settings (Full Model):

Steps: 22 - 30
CFG: 4 - 10
Sampler: DPM++ 2S a / DPM++ 2M SDE
Scheduler: SGM Uniform

Settings (Hyper 8step):

Steps: 8
CFG: 4 - 7
Sampler: DPM++ 2S a / DPM++ 2M SDE
Scheduler: SGM Uniform