If you like my work, drop a 5 review and hit the heart icon. it's free and keeps me motivated
Free GPU use on Tensor.art
Mage provides Unlimited Generations
Made a new Patreon account. do check it out
I post my upcoming concepts and nsfw lora early on my patreon
If you want commission dm me on discord.
Get early access to my upcoming NSFW Lora in my Patreon .
Support my work by joining any one of them and get early access to all my upcoming loras and other perks such as fan requests and Discord role.
Suggested settings:
I had CLIP skip 1 on every image
Samplers : Eular A, DPM++ 2M
I had ENSD: 31337 all of them
All of them had highres.fix or img2img at higher resolution.
I don't use restore faces
I use afterdetailer for face details
Tiled diffusion for img2img upscaling and sometimes Noise Inversion for some more extra details.
4x-Ultrasharp upscaler
Join my Discord Server
Description
FAQ
Comments (20)
loved version 1 , whats the difference now with the new v2? might just instantly download it
The second version solved the frequent portrait painting issues in V1
The second version is not as good as the first version, but the first version is prone to portraiture
V2 no longer usable on Civit?
When someone said where's V2, it's back and V1 is gone... Ah, I love how ironic it is...
I have used both versions, and I think V2 is better. Compared to V1, it is softer and more exact,and it fix the portrait issues.There is only one version can online, if given the choice, I would choose V2.
First off, I want to say: I absolutely adore this model. Hands down, without question FAVORITE model. Goofy_Ai created something amazing with this.
Long Review Below:
Since release I'm not even going to say how many images I've pumped through this... The answer is: Yes. This has replaced my curiosity for Perfect Pony v4, and retired Perfect Pony v3. Yes, these are TWO drastically different models, and create different images. But this style I prefer WAY more, personally. However, I will not deny after more testing, Perfect Pony V4 can crank out some absolute epic anime style images. For me, I wanted something more, without knowing I wanted something more.
After some tweaking and playing with all the different settings I found my home that will generate amazing images. A Lot of plotting went into that, and having to study the fine details to see what changed.
The ability to create vast and drastically different characters is awesome. With some testing on the Anthro type characters it can create sharks, wolves, dragons, cats, very easily. No LORAs needed. Instant results with a quick prompt.
Speaking of LORAs, this is the down fall of this model. Actually it's the downfall of semi realism. Most all LORAs trash the quality of the image. Since most are designed to be used for actual anime style. Then going to realism LORAs, those trash the results making things too realistic. The other, is it just overrides the character prompt enough or changes how it generates light and saturation, it becomes two different characters. So for me, LORAs have to be very picky to use. Right now I resort to seeing if I can get the model to generate it or get close.
At this point, I feel like I've tested everything I could some of the most random things I could think of, it handled it pretty well.
I'm not sure if it's because I tested Pregnancy more in Sym-Pony World, or didn't test enough in Perfect Pony V4, this seems to have a heavy, heavy training bias on Maternity photos would be my guess. Their for, it forces most to always show a nude stomach, or tight pulled clothing around the belly. I feel like I was fighting that far more than the tests I did on the other model. Also the weight to hold the belly is insane. Which again, makes be believe maternity photos were a core of training.
While I am on the subject of pregnancy, this will tie into the next subject breasts. I did notice a direct link to belly and breast size as well. Example: "Small breasts + Pregnant belly", usually doesn't look pregnant by much, or very sparsely looks right. I was noticing around "(Pregnant belly:1.3-1.6)" before the subject would look 8-9 months. But "large breasts + pregnant belly" would result in a much more pregnant version with no adjustments. This sent me down a deeper rabbit hole out of curiosity, but that isn't for this. TLDR: I think we need very well trained LORA or bake into model for pregnancy content that actually recognizes very common terms. ("8-40" weeks, "2-9" months, Pregnant with twins+, early pregnancy, "X" Trimester, etc.) Personal vote would be bake into model so you won't change your characters with it.
I won't really call this a flaw, but something I was testing when creating a character for a story. Both models need better training data on small to huge breasted females. 98% of the images I generated with the "rare" exception of a few, when using small breasts prompts will generate a MUCH younger female, just as large/huge breasts ("massive" doesn't work in sym-pony) generates a much older female. Why can't I have a hot tomboy female, that doesn't look 14 years old? Because training data, that's why. I really do ask Goofy_Ai to look into this when making V5 for Perfect Pony or V3 of Sym-Pony. I would love to see this fixed. For reference:
Pos:
(mature female:1.6), MILF, (tomboy:1.6)
Neg:
child, young teen, young adult, teen, (petite female)
With "small breasts", in the prompt it will always still result in a younger female, no matter what. You would think that would look like "Stacy's Mom" but, small... NOPE. It was Stacy. EVERY time...
One other thing I noticed here was backgrounds are usually good. There are few that are VERY bland and may need more data. Malls, shopping or any kind of shopping/ markets usually results in like 1-5 image backgrounds. Can't make department stores, or good farmer's markets. So just as food for thought, maybe look at background/scene improvements someday.
Another big compliment I can give this model, with around 30-40 steps and 15-20 on hires fix, you generally could skip ADetailer on faces. However, the touch up always looks amazing, and usually more refined. But the fact stands, it's 100% useable without.
Sexual content is pretty good. But there are times where maybe it was the prompt or concept that made it not want to put the male into the shot. Example: "Doggy style" with a few other prompts will get a girl on her knees, and a male behind her. I do end up seeing just the female a lot with that test. I assume it's just the prompt since it does appear other characters ideas have easier time than others. So not 100% on model or not. But again food for thought, and something to look into.
I cannot again say how much you will enjoy this model. 10/10 would recommend. My advice for any who experience bad image generation, it's your LORAs, you have TOO many on your image. That's why I looks terrible. Try to NOT use the LORA unless absolutely necessary. The raw model will generate absolutely stunning images, but your LORAs are mudding them up.
Goofy_Ai, props to you and what you do. You have my thanks and gratitude for your efforts into your AI models. You've made me a true believer of your work and you have a fan out of me with your models. I'll never go anywhere else for my models.
thanks for this long review. ❤️
some of your questions answers are yes there is a training bias for woman as its the case for whole Danbooru site. and you are also correct about the lora and semi real problem. you might have some lunch with adding realistic in your prompt but then again most of the loras are trained on pure anime images so they will impact the model heavily.
finally a piece of good news if anyone cares. a new middle ground checkpoint of prefect pony and sym-pony model based on illustrious is very close to final rollout.
@Goofy_Ai You are welcome. It's been a lot of fun playing with the models. Yeah, I've been playing with "realistic" on things like Anthro (X) and that sometimes brings it back to more what I'm seeing. Then other's it's just too real for what I'm after. But seriously thank you for your work! I know the community appreciates it even if they don't say.
I take it this Checkpoint isn't the one currently up for illustrious (Prefectious XL NSFW)? I've been testing that model for the last few days. Was looking to do a write up on it as well. So if a newer / better / different / updated version is coming, I'll hold off on testing further.
@Night1176 different from that one. Prefectious is anime style. You can check the latest character Lora I posted. The kaguya one. They where with that model I'm talking about.
@Goofy_Ai Ok cool, you have my attention! I look forward to trying it out and giving it a full and honest review! Looks promising.
I LOVE this model though im a bit bummed that v1 is not available in civitai anymore but i do recognize that its not the fault of the creator but the site itself...
I know its available on other sites but i really prfer civit in terms of ui and just feel so ill just mourn the loss by my lonesome :(
You are not alone, I'm also sad that V1 isn't available here anymore and I don't even know the reason. I just know I had started working on a whole set of pictures back in November which I tried to finish these days, but gave up when I noticed how vastly different V2 looks. The inconsistency within the set made me lose interest, the whole effort seemed pointless and, of course, thanks to Civitai's overall generation problems during the past two-three days, the last batches of those pictures ended up being auto-deleted before I could do anything with them, so a whole lot of time, effort and Buzz went down the drain, just like that, for nothing.
The exact same thing happened with the 2DN checkpoint a couple of months back, so I really need to ask - what is the reason Civitai disables older versions of checkpoints when new ones are released? It's annoying, to say the least.
@MakiZeninHasNoScars Though im not 100% sure that this is the case for this specific model, Civit usually limits popular models that are available on other sites if they don't have exclusive access.
The creator of another popular model, Nova Anime XL, has said that the reason their model isn't available to be run in civit is due to the fact that Civit wants them to only host the model on their site and not other sites such as Tensor or PixAi.
@Kimlas Thank you for the explanation. I had no idea hosting and maintaining models was so bothersome. Not the best thing for your average everyday amateur user like myself, but what can you do.
@MakiZeninHasNoScars Organize a protest to stop the anticonsumer practices. Civitai is trying to use its position to bully users and creators alike to push them away from other sites. Everyone has to get together and show Civitai that they will not put up with it. If Civitai wants to play this game, then go to the competition where the checkpoints are available. We need competition.
If you do nothing, then nothing changes, and Civitai continues down this track and just keeps getting worse.
This hands down the BEST Pony checkpoint I've ever used.
I love this model and it's 2.5D look. Maybe you can sometime make it work as good with loras as Prefect Pony? Keep up the great work.