NOTE: This model is a hybrid mix of SDXL and PONY bases. Unfortunately, this isn't officially documented, as I started it over 6 months ago when I didn't have much experience. Because of that, I can't credit all the specific components involved—there are probably at least 7 models mixed together, and even more if you count some of the other people's mixes I used. To everyone involved, thank you so much for your hard work! In the end, this experiment resulted in a model I’ve been using personally for all that time. 😊 I love this community and wanted to share it because I believe people can make some truly great stuff with it, just like I have!
To the point: This model is an overall well-rounded one, in my opinion. It can handle a variety of tasks, including:
SFW & NSFW content
Beasts, characters, abstracts, landscapes, and complex scenes
Perceived physics for skin and liquids
Excellent at image-to-image transformations (e.g., frame-by-frame animation sets)
If you're familiar with Gelbooru, this model works well with the image tag list—about 70-80% of tags work fairly well. The model also shines with LORA. I generally run 5 LORAs at once: 2 are purely scalers/enhancers, and 3 are style or character LORAs. This model is compatible with SDXL and PONY models, and with the right prompting, you can achieve near pitch-black or extreme saturation scenes.
Every time I think I’ve seen all it can do, I discover something new, and I end up using it all over again! 😄 Below is the layout of the settings I usually work with.
Settings I Use:
Clip Skip: 2 or 1
VAE: SDXL_VAE
Sampler: Euler Ancestral
Scheduler: Simple
Resolution Size Variations: 1024x1024, 768x1024, 1024x768, 1216x1216, 832x1216, 1216x832, 1200x1200, 1200x1744, 1744x1200
Inference Steps Variations: 25/40/65
Guidance Scale Variations: 4/6/7
Upscaler Variations: RealESRGAN_x4plus, RealESRGAN_x4plus_anime_6B, Nearest, Lanczos
For Pony Prompting:
I generally stick to the basic tags:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, rating_explicit
These are not required for the model to work, but without them, you’ll get more of an SDXL vibe. If you add them, the result will feel more like a hybrid between SDXL and PONY.
Embeddings I Use:
Positive: fixl-art_
Negative: zPDXL-neg, NEG-fixl-2
Both fixl-art_ and NEG-fixl-2 are triggered by the same word, just in positive or negative slots. These, along with everything else mentioned here, are available for download on CivitAI. They have no association with me, I’m just a user of them.
Enhancer I Use:
The only detail enhancer I use frequently is add-detail-xl with a strength setting of 1. This, combined with higher base resolutions, produces decent results even on my rig (which isn’t great). I get generation times of around 25-45 seconds per image.
Basic Setup Suggestions:
Inference Steps: 25
Guidance Scale: 4
Base Resolution: 832x1216, 1216x832, 1216x1216
Use 832x1216 for full-frame or cowboy-shot style characters, 1216x832 for lying-down poses or landscapes, and 1216x1216 for group shots, complex high-standing scenes, or close-range sexual posing.
Model Behavior:
Without any influence, this model tends to generate about 80% adult-based characters with a standard adult physique. It can handle nearly anything I’ve tried with it, including size differences, furry beasts, and extreme sexual content. However, it struggles with specialized fetishes or extreme gore (like entrails or viscera) without the help of LORA.
General Prompting:
This model works well with simple prompt structures, like basic tagging, but also supports more free-form descriptions, such as music lyrics or paragraphs from books. If something isn’t working, just try swapping a few words out and generating again.
Final Thoughts:
A huge thanks again to everyone who made mixes like this possible! I hope you enjoy this model as much as I have. Feel free to post anything you create with it. I don’t believe in censorship, so go crazy! If something gets removed, it’s because CivitAI didn’t like it, not me. 😜