I rose from the dead so to begin with:
V9 - better at everything but a little less natural and more “AI-ish”, doesn't need negative embeddings and detailer every time
V8 - maybe inferior sometimes and trained too much so deviate from the “pony stuff” but believe me it makes a great base for teaching lora for specific characters etc
Tips:
Use ADetailer and/or HiResFix!
Recommended CFG ~5 (7 is ok but less realistic)
In ADetailer recommended denoising strength ~0.15 (0.1-0.2 is good)
In ADetailer to avoid some kind of cutoffs "Use separate VAE" and set it to standard sdxl vae
To avoid oversaturated images when using HighResFix change vae during process to standard sdxl (there is a separate button in Automatic1111 webui to create an upscaled version of the currently generated image)
For X and higher models recommended sampler is DPM++ 2S a
If you like my work you can tip me on - patreon.com/Natakaro or buymeacoffee.com/natakaro :)
Description
FAQ
Comments (15)
Any recommendations for comfyUI equivalent of ADetailer?
DDetailer but I use comfyUI occasionally so I don't know if it works well for sure
FaceDetailer from Impact Pack
wooooow ! just the best realistic pony model *****
can you explain how you managed this exploit ?
Thanks :) I trained step by step loras (with small rank and alpha to not "break pony stuff") - train lora on small dataset, merge to a model, find issues, create need small dataset that could correct issues aaaaaand over and over again.
PS I am preparing a fix, the current model breaks down when you add something about eyes or use some facial expressions eg. green eyes - eyes are getting bigger and anime-like
@OsorubeshiAI really good job 👍, you are skilled ! I use your checkpoint in my merge, in fact the two only reason i merged it are : cause your version is a little bit unsaturated , and second his more personnal i find it not enought style versatile (i used to use versatile model and after it use lora when needed).
Are you planning to realese a realistic Lora for base pony ? (In this case it would be really interessting, giving us the ability to control the amount of "realism")
@Qentah Yeah working on it but it gives some artifacts when using under 0.9 strenght. For saturiation use xlVAEC_f1 vae ;)
by FAR the most realistic pony of any available. There are a few abnormalities that pop up, but if you are batching, at least 75% will be decent. This pony also seems to handle both pony and straight SDXL loras well. I have been able to get some nice shadows and lighting, while still getting good results, which rarely happens with other ponys. Well done!
If only there wasn't a different person on the page of every single realistic Pony claiming that IT'S in fact the best one lol
@diffusionfanatic1173 Yeah XD It's quite funny, but it seems that everyone generates images in a different way, uses different tools and prompts - so everyone gets different effects and likes different stuff, each model shines under certain specific conditions. The best results honestly on pony can be achieved by training lora on one (and only one) specific realistic character.
I'm a huge fan of your v2. I guess my main question would be, what do you see as the updates and changes that went into v3. Limited disk space forces me to triage my checkpoints. And not sure I can have two of yours at the same time.
v3 is more realistic, more photo-like less "AI" bloomy and blurry, but also I think it is a bit harder to use
PS there will probably be a v3.1 or v4 in the coming days
@OsorubeshiAI Great. Thanks for the response.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.