— v5 PONY Update —
Improved Pony handling
Fine-tuned prompt sensitivity
Bug Fixes
Better Textures
— v4-SS PONY Update —
Better prompt handling
Turned down some sensitivities
Bug Fixes
Better fantasy knowledge (not perfect)
— v3.3-SS PONY Update —
Better anatomy and human textures
Reduced need for complex negatives
Increased prompt sensitivity
Reduced need for quality prompting
READ UPDATE FOR NOTES ON USE
SS = Super Sensitive (I'm working on a less sensitive release of 3.3 as well)
NOTE
This version is very sensitive to quality prompt tokens (saturation, ultra-rez, stuff like that) If you get a solid color generation or noise generation it’s likely because your prompt is over-baking quality or you’re using an overbaked Lora. Try removing that token or reducing weight (Lora weight reduction from 1.0 to 0.8 usually solves it but you may have to go lower if the Lora is super-baked)
This is still doing well at 30-50 steps. Once I dial the quality in where I’d like it, I’ll work on reducing required steps to speed things up.
*I've had two people mention that they were having problems with this model, I've asked them to share any workflows and prompts so I can look into it. If it is a problem with the model, I'll issue a hotfix for v3.3x once I solve it. Other than that, remember that this one is pretty sensitive and may not need super elaborate prompts.
— v2.12 PONY Update —
Better Prompt Adherence
Improved Textures
Realistic Texture and Lighting improvements
** I'm still learning Pony training, so bear with me as I work out kinks in my system. This model is still a little bit of a bugger when it comes to animals. I've improved aging, but you may need to beat it over the head a bit in your prompts: e.g. Elderly man/woman, old man/woman, gray hair, wrinkles etc... The model definitely has a hotness problem that I'm working on (and making progress with). I've made some changes in the training to get it to produce different faces more frequently - I've noticed a lot of Pony models seem to be trained on specific faces so I'm trying to break it out of that a bit. It is pretty good at maintain a "character" within a prompt. So the person you get on gen1 is usually the person you get on gen20. Anyway, it's still pretty good between 20-50 steps and I've tried hard to make score prompting work but optional. So you'll see variations depending on whether you use that or not. Other than that, enjoy and let me know about all the cool problems you discover so I can work on sorting them out in the next version.
— v1.1b PONY Update —
I’ve been asked several times to drop a pony version, so I have. BUT… this is my first Pony model and I’m learning a lot about how to tune this sucker so bear with me. This version is pretty good with people but needs a bit more arm-twisting with things and animal primary subjects. I’m also testing early access as a way to help explore some future FLUX versions. Anyway, enjoy and let me know what I can do better or even if you just like it. It responds well to 20-50 steps and score prompting works but isn’t necessary.
-- v3.1 Update --
A couple of bug fixes
Better upscaling response
Improved Skin/Hair textures
Better prompt adherence
-- v2.5 Update --
Better Prompt Handling
Improved Textures
Improved Lighting
-- v1.11a --
My models always start as an update to an existing model until I do something weird.
Anyway, while training this one I discovered something that had a huge influence on my training data and refinement. I'll treat that as the secret sauce for the moment, but it ended up diverting away from my initial plans and had to become its own thing.
So this started with my model Aardvark 2024 and then quickly started going down a different road. About half-way through this I was completely blown away by Pony and I have to give so much praise to @PurpleSmartAI for that project. Really amazing. This model does not include Pony at all though, I was just inspired by the work and it changed some of my thinking.
This is definitely starting as a v1.0 - and I will be updating it, more quickly I think due to how everything came together.
Let me know what you think, and I'll incorporate your feedback into my future training.
Oh... I get asked a lot about whether I'll turn these into SDXL models - I'm not sure yet and I have a pretty good reason. I train and use embeddings a lot for my own work and SDXL has been a bit of a bugger for me to work around. I'll keep experimenting there but I'm not going to release anything until I know I have it right.
I want to add a shout-out to 3mcg33 who has some incredible prompting skills. I frequently try their prompts when I'm testing models and I want to make sure I give credit where credit is due. The chimp in the suit is a great example, 100% inspired by 3mcg33.
Description
FAQ
Comments (10)
he did the impossible. he removed the valley sized clefts that plague every single realistic pony model.
"My models always start as an update to an existing model"
Isn't that kinda overstating the obvious? Since that's what pretty much all SD/SDXL/Pony models are, with exception of the first few created?
No, he doesn't have to continue from previous iterations. V2 doesn't have to have anything similar to V1 apart from the SDXL base model. He is implying that he is also including training from V1 in V2, etc.
What @makiaevelio543 is saying is accurate but also, my models typically begin as an effort to work on one of my other models. With Arrogant Bastard, I was actually working on an update to Aardvark when I took a direction that deviated from that foundation. So Arrogant Bastard really no longer fit with that model and I felt like it should become it's own thing.
I guess my real point was that I only ever set out to create one model here for my use and everything else I have done has been the result of a deviation in training a version of that model where the output significantly stepped away from what I was working on and became it's own thing.
Maybe I could have done a better job of stating that.
You are right in that all models are deviations on the original SD/SDXL - Pony is just a fine-tuned version of SDXL.
In this case, I was looking to update Aardvark and during that process I discovered a different method of training some assets... and that fundamentally changed where it was heading.
@QuadPipe To be fair, I did pretty much get the jist of it. I was just being a bit of a smartass due to the phrasing. But also being a bit informative, cuz a lot of people don't realize that most of the mainstream models are effectively just alterations of only a few. I'm not sure I understand the name of your model though. 'Arrogant bastard'? If someone didn't know better, they may think your model just produces images of that guy in the sample shots. I'm sure most won't, and those that do will just move along and look for something else. Though, names that emphasise what the model actually does, is always a good selling point.
@Lazman The Arrogant Bastard name was just me poking fun at something someone said to me here once and I was being a bit snarky. Also, I liked it because it started with "A" and was at the top of my models list when I was in A1111 or Forge.
@QuadPipe I shamelessly do the same things with my local checkpoint merges, _, 000, a few others lol
@makiaevelio543 Just so long as you don't distribute models for others to download with the full name being 3523524.safetensors .. I got a bad habit of downloading a buncha models then going to check on them later, and when they're named like that, it's like.. WTF is this thing.. There's some people that I fkin love for their model naming though. Like, I go to check and it's got the model type, descriptive name, and even a keyword or two in the name. It's the best practice imo. But at least to have the model type if nothing else. Makes it easier to test the models/loras to find out what they are, even if the name isn't descriptive.
I've got some models I never got around to testing cuz I don't know what they are, and the odd time I've tried to test them I get errors about them not working with my other stuff.
I've been more on it recently though, with adding the model type to the beginning of the name of every model/Lora I download. I've been using comfyui recently, and it's not the easiest to sort the model types otherwise. At least not that I've found yet. Only been on it a few days now. I'm sure there's some extension/node to make things easier.
can't do futa
The best realistic pony model I've seen (and I've seen a lot of them and only one of them can be of such great quality).
212 copes with realism perfectly. But 33SS seems to move away from realism. Generations on version 33SS have a certain cartoonishness, so for now I have settled on 212. Thanks for such a good job. You're doing great.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.