@epinikion I was extensively testing your model, and for everything photorealistic - it is amazing, but for multitude of NSFW poses "normal" SDXL models are just not the right tool

yeap @OliviaRossi is 110% right! Pony is not only NSFW (but, for the love of god, Pony is the best in the human anatomy) but it´s the best on prompt interpretation and that is because it was training on booru tags. I'm a big fan of all the pony models and I reflected it on my youtube channel, now the community only use pony hahaha

@stable_diffusion_espanol agree, that Pony method of training is superior for all future models, since we as humans want to see interaction and posing mostly of humans in images

@epinikion there is a small but also important detail - Pony was trained on semi-realistic/cartoonish images = it looks that this is the reason why it is able to "draw" human hands and fingers way better (not only overall pose), and as soon as we push Pony to realistic generation that ability diminishes visibly

@OliviaRossi The prompt responsiveness what the first thing that impressed me the most with Pony, but I still see so many post where the image is undoubtedly incredible, yet when you look at the prompt, it doesn't line up. Seems like most rely on loras to shape the generation, I've been searching for a pony model that ultimately does what you want before adding a bunch of loras. I've always loved @epinikion models, your work is amazing.

@DigitalAF LORAs are only necessary for Pony when it is blind (untrained) for some NSFW stuff like fisting or some particular sex positions or outfits, but mostly LORAs are only used for style (and that is the same as with regular SDXL models)

@epinikion Pony needs a specific way of prompting (i.e. Danbooru tags) to see its benefits. A typical prompt that works best with Pony looks like "score_9, score_8_up, score_7_up, BREAK, [a short natural language description], [list of Danbooru tags like '1girl', 'blonde_hair' and so on]".

There's no denying that Pony models are best for NSFW content, but I recommend experimenting with prompting styles if you haven't already.

I use this site to search the tags and their respective popularity before including in my prompt, by the way:
https://nsk.sh/tools/danbooru-tags-explorer

@epinikion Just replace the Unet of any Diffusers model.

https://huggingface.co/cosmicman/CosmicMan-SDXL/tree/main

@epinikion 
# 1

# checkpoint to diffusers

# from https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py

python3 convert_original_stable_diffusion_to_diffusers.py --extract_ema --from_safetensors --to_safetensors --image_size 1024 --original_config_file sd_xl_base.yaml --pipeline_class_name StableDiffusionXLPipeline --checkpoint_path epiCRealismXL.safetensors --dump_path CosmicMan

# 2

# Replace unet, diffusion_pytorch_model.safetensors and config.json

# from https://huggingface.co/cosmicman/CosmicMan-SDXL/tree/main

# 3

# diffusers to checkpoint

# form https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_sdxl.py

python3 convert_diffusers_to_original_sdxl.py --use_safetensors --model_path CosmicMan --checkpoint_path CosmicMan.safetensors

@andycoco thanks for sharing!

I agree with you - I changed your PROMPT a little and am speechless with enthusiasm!

Results look great. So you're saying switch to CyberV50 for the hires fix step? What sort of Hires prompt are you using and How many steps and denoise?

@arnorwing 

Upscaler: 4x_NMKD-Superscale-SP_178000_G Upscale 1,5

Hires Sampler: DPM++ 2M SDE Karras

Hire steps 100

Denoising 0.26

pos: raw photo, {prompt}, uhd, 8k, film grain,

neg: (deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation, (artist name, signature, watermark:1.4), (ugly:1.2), (worst quality, poor details:1.4),

@slartibart353 100 hires steps? That's just a waste of time.

@Aptronym yes I agree. I currently use 6 steps .3 denoise, in my testing of upto 20steps there is no worthwhile difference. 100 steps would take ~10minutes on a 3090.

@epinikion agreed... 4090 here - 45 secs ;-) but 15 steps does not work for me - lost details when usind SD1.5 hires.fix