As this is a merge based off PonyDiffusion v6 I HIGHLY recommend you read the description of that model for information regarding prompting, tags, and baked in characters.

TL;DR Generation Settings:

• Sampler - DPM++ 2M Karras / DPM++ 2s a / DPM++ 2s a Karras

• Sampling Steps - 25

• I upscale to 4x using this process and the upscaler below:

  • 4x_foolhardy_Remacri

• Dimensions - 832x1216, 1216x832, or 1024x1024

• CFG Scale - 7 or higher

• One or more passes of ADetailer

  • Some combination of the following for faces and eyes

    • Facesv1.2

    • Anzhc's Face Seg Prototype

    • Anzhc's BeautifEYEr

I use the following positive and negative prompt. I DO NOT USE ANY SOURCE TAGS (source_anime, source_furry etc) :

• score_9, score_8_up, score_7_up, BREAK {prompt}

• NO NEGATIVE PROMPT OR EMBEDDINGS REQUIRED

  • simply prompt for what you don't want to see as it comes up

See sample images for examples.

Heavily stylized LoRa merge, tweaked to my personal taste. Has compatibility with most concept/pose/character models, but hands are still pretty hit or miss. Models I used are listed in suggested resources.

As with most of my mixes, this is intended to produce female characters and leans towards NSFW.

V4:

Fourth time's the charm. I wasn't happy with the outputs of V3, so I baked models of my own and remerged to get V4. This will be the last(tm) update, as I feel it's stylistically close to what I wanted when I first made the merge. Watermarks should not appear by default (I encountered none when generating the samples).

V3:

Rebaked an underlying LoRA that results in the majority of the watermarks introduced in V2 being removed; no more needing to prompt for things like watermark, patreon username, artist name, signature in the negative by default. Otherwise outputs should be very similar to V2.

V2:

Remerged from the bottom-up using modified weights. Less stylistically intense (especially around the faces and eyes), more inline with what I had originally intended for this model, but still not quite there. Reduced the need to use specific prompts to achieve the desired look. Hands are still hit or miss, and there's a persistent watermark that appears ~50% of the time: some inpainting might be required. Hoping to rectify this in a future update.

The baked-in artist styles for PD should work, but as this is effectively a style merge, they will defeat the purpose of this model.