This model is a merge of BigASPV2 and older model called onlyfornsfw that didn't see great success but I like for it's ability to do nsfw stuff very well.
I run GGUF of the model to fit into 6Gb VRAM.
For settings I recommend DPMPP 2M with exponential scheduler. PAG with scale 2-3 raises output quality a lot but takes slightly longer. 25 steps and 3.5 CFG seems to work well.
Description
FAQ
Comments (5)
Thank you, Great model.
yeah, onlyfornsfw was the best model of the early SDXL times. sadly it disappeared. Honour to the creator
Do I need BigASPV2 clip L and clip G to run with the GUFF Q8 checkpoint? I ask because there are default clipL and clipG files already which I've taken from huggingFace.
If they are the default SDXL clips, then there's a significant difference. You can of course try first.
Great! I generated some pics with your ClipL and ClipG and I uploaded a sample in the review. My method: 1) I generate a 1216x832 with {LORA} LCM_TurboMix_8step_CFG1,5_2,5_ponyxl_illustrious_SDXL_EulerA , with EULER A, EXPONENTIAL, 8 steps, CFG 2,5, without specifing details like genitalia, mouth, fingers, etc (in up to 50 seconds per generated image). In the 'refine' phase (UPSAMPLE), I do a 1.8x (2184x1496) with "CR UPSCALE IMAGE", using 4x-Ultrasharp, then I pass this to a second KSAMPLER with DPM_2M EXPONENTIAL, CFG 3,5, 19 steps (up to 7 minute generation in a 6GB NVIDIA card). Of couse I can just generate with LCM sampler (say, 20 images), then I select just those I want and BYPASS KSAMPLER1, activating only KSAMPLER2, with custom positive prompt with the details I want.