- Added trained Lora

- Cleaned style

- Baked in VAE

- Extra trained data of 15000 steps

- Better Lora support.

@AshWasTaken I made a comparison post, it will appear soon.

@NMackelie remind us when the article appears for your comparison post. I am also interested in reading 👍

@AshWasTaken @ronaldmikhailp236 https://civitai.com/images/39390336

no

@zxy123333 It seems like this is an issue with the new v4 model. I found an ADetailer workaround, but it should work out of the box. I think the model has a lot of promise, especially if you like to have good backgrounds, but it probably needs a 4.1 as soon as possible to fix the current issues.

I noticed that too. ;-(

All the previous lora's made with version 1 and 2 look very different with 3 or 4 now, not like intended anymore.

I checked many lora's I liked before, that were made with ver 1 and not a single one looks as nice and like intended with the newer versions.

you should just keep V2 and V3

Hands look fine, toes look fine. Double limbs are almost nonexistent. I have a couple of weird hand positions and the eyes are the biggest issue at the moment. Be it empty eyes for Hypno/mc, constricted pupils for the shock effect, or straight-up normal eyes, they look weird. ADetailer, which I use often, makes things even worse in some situations, creating multiple noses if I don't reign it in and messing up eyes even more. I hope this gets fixed because the Backgrounds are one of the best ones I have seen so far since I started back in 2022. I am especially impressed with the consistency and tables/windows, even if interrupted by a person in front, don't suddenly shift into a different position when they reappear.

I don't understand why the author has removed V2.

It was my Favourite.

@LoraCraft Same, used it few days, but this was better than AutismMix.

@HotitiDracoony it was awesome for anime and cartoon stuff

@Goofy_Ai oh ok was going to have a heart attack, you know why there is a limit.

@Goofy_Ai Limit for what?

@Goofy_Ai will V2 come back?

I think this is the Best idea.

x2 v2 y v3 good

Yeah, V4 is Very similar to V3 but V2 is great for Anime/Cartoon Stuff.

I wish the Creator would just switch V3 for V2.

yes it is.

Goofy_ai just needs to swap V3 for V2

I agree boop

@blazerhuge42 my life is ruining without V2

Nope, v2 it's still better.

i tried v4 and i get better lora usage on v3

keep v3, remove v4, v4 literally has no big difference from v3

@kuraig if v4 literally has no big difference from v3 then replacing v3 with v4 shouldn't be a problem no?

@Nightcliff I didn't get the chance to use v4 with Loras, but using it on its own gave me much clearer and higher quality images, especially for sfw prompts.

Naw, V3 and V2 are best.

@69stormcloak69903 the boob and face look weird in v4, that's why I prefer v3, I've tested it with x/y/z plot

More like keep V2 only Ngl, or just make the V4 cleaned version like what V2 does

@Goofy_Ai Ok, I'll try

https://civitai.com/posts/9015964

i teste it with same seed same loras etc. and its looks worse. it makes for me way to thicc outlines and is less smoothe. so i stay with v3

wdym draw thing?

@kuraig it’s an iOS app that lets you run stable diffusion locally on iPhone or iPad (and Mac)

i recommend mixing fffmix (available here) + prefect3 or 4 at 70/30 fffmix/prefect. Gives the desired result you want probably

I dont like a anime, but love ponies. V3 and V4 doesn't do them very well

The site has dropped support for V2 either by deploying it yourself or by switching to another site that generates it online.！！

X2

V2 the best for pony from My Little Pony

V2 is the best, it's the model I've used for most of my pictures, but now I have to force myself to use V4 and V3, and I personally don't really love any of them.

I tried V4 and it looks ugly. Watched someone use V2 and it's way better than V4

Try things like:

Pos Prompt:

(blonde hair male), male with blonde hair,

Neg Prompt:

Orange hair male, Male with orange hair,

I feel your issue might just be a technicality. Understanding prompting and how to massage things out more efficiently. I looked at your three images and noticed a few things. For those being Perfect Pony V3 and 4, they aren't very sharp even for 28 steps on Euler A. ADetailer FaceYolo would do wonders for your faces and eyes. The Lora was designed for Clip Skip 1, but the model is meant for Clip skip 2 from what I've tested. Even Goofy_Ai says in the description Clip Skip 2 was used on every image. So that can cause some issues.


"score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, (RAW photo, best quality), masterpiece, extremely detailed, highres, cinematic lighting, best quality, High definition, 4K, detailed, detailed body, volumetric lighting, vibrant colors"

^ all this crap is probably ruining the model on V3 and V4. This isn't Pony Diffusion, it's a merged model of Pony with MANY Lora's at this point.

detailed body, volumetric lighting, cinematic lighting, vibrant colors

^ try at the end of the prompt. The most important info needs to be first in the prompt: AKA, Your character details. All the Lighting and LORA specific can be at the end.


The user doesn't specific if you should use the LORA on "1" or "0.X", but common logic says: if you want the model to do more guess work: 0.7 or less could work. If you want the LORA to do more guess work "1" is more ideal.

If the issue is detail, try to find a Digimon style LORA. Then use that to get the "look" of the show. The issue isn't the model, in your case, I'm 90% it's how you are using it. For how the details are smeared, this feels like prompts are conflicting. That image "should" look far better, not like a cheap Instagram Anime filter.

(Sorry if it feels like I'm bashing, I really am trying to be helpful and hope some of this advice might help make a better image.)

https://civitai.com/posts/9431012

Here - I put my money where my mouth was. A Perfect Pony Version 4 Gen with Anime specific LORAs.

@Night1176  The "cheap instagram anime filter" part, lol 😭

But seriously, thanks for all the valuable advice, I'll definitely look more into it and do some testing! Like I said, I just do this for fun and, to be honest, I've just been winging it since the beginning, started by copying people's prompts and kind of just keeping what seemed to work best for me... 😅

I have a couple of questions, though, if you don't mind me asking...
1) When you say "all this crap is probably ruining the model on V3 and V4. This isn't Pony Diffusion, it's a merged model of Pony with MANY Lora's at this point."... Should I ditch all that "score_9, score_8_up, etc." in my prompts for this model, then?

2) Could you clarify this part: "if you want the model to do more guess work: 0.7 or less could work. If you want the LORA to do more guess work "1" is more ideal"?

Sorry for the noob questions and once again thank you for the advice! :)

@randomuser_mon xD yeah I mean it's the unfortunate truth. looking at he LORAs they should have looked like the anime, not a filter. But the idea is the help you, not bring you down.

1) Yes, stop using scores and quality tags in Merged Pony Models, especially Perfect Pony. So on the example I posted for you, I didn't sure a single score, "image" quality prompt. I focused on what do I want my image to look. I wanted the checkpoint and LORAs to do the guessing. I knew Perfect Pony could create amazing images without the need for any of that. So try your images again with out all the: Score_9, High Quality, etc, prompts. - Now that being said, saying something like "high quality hands, detailed dress design, or beautiful eyes" as prompting, this is actually ok. These tags aren't the same. It just tells the checkpoint, to look for what the creator said had those descriptions in it. Just know, sometimes, this backfires. Lastly, look at using ADetailer and FaceYolov8n. This will make your faces far better detailed, and usually makes eyes look 1,000x times better.

2) So prompt weighting comes like this:

- <"X lora":1> is usually how it looks. This could be automatic1111, Forge, or in ComfyUI it's a parameter. On site-generators it's a parameter most likely.

First understand what this means

"1" - Is default, this means it's going to look at the LORA with full weight.

"0.7" - is saying, use this LORA, but let the model do some of the guess work. Meaning this will use the LORA, but it won't impact the image as much. Say you want a 2B outfit on your OC. "1" might make her look more like 2B, where "0.7" might stay looking like your OC.

"1.3" - is saying, focus more on this LORA, What I need is from here. Focus on this.

This is also the same for Parameters, but use cautiously. Example: "Chubby Thighs" is a prompt you want to use. All you want is a little thickness, nothing crazy. The girls legs are still skinny. Try: "Chubby Thighs:1.1" or another idea: "(Chubby Thighs)". This will tell the model to look at this keyword a little more and add a little more weight to the concept. If that doesn't work try a few more. "Chubby Thighs:1.2" or "((Chubby Thighs))".

LORA in short is a small model checkpoint, Hyper focused on ONE character or idea. Where the Model Checkpoint is focused on the overall quality of an image. So basically, look at the weights as a control leveler.

That being said, If you actually read authors pages for LORAs, they will tell you, normally, how the LORA is intended to use. "Best results use: XYZ" Doesn't mean it won't work if you use a different number, but it may not function the same. In your case, they didn't say. For the Kirito images I made, I use Kirito's at 0.7, and the SAO style at full value 1. Why? I wanted the Perfect Pony to do some of the work on cleaning up his character and poses, but I wanted the image to "LOOK" like it was from SAO. Each LORA is different, you just need to experiment with it.

Then lastly just as food for thought, try other LORAs of the same character, some might be made better. Yielding better results. I've seen some amazing and terrible made LORAs. Just because it's downloaded it lot, doesn't mean it's correct. And the other advice, make sure you use Pony LORAs. SO far I've had mixed results using SD LORAs. Not saying it doesn't work, or come out right. But so far, Pony LORAs work far better.



Another resource for you - Try these in as your #1, and #2 Negative prompts. These can also massively change, but massively help your images. So just experiment and see what works for what you want to see. Me, I run both. Both? Both. Both is good.

https://civitai.com/models/332646?modelVersionId=720175

https://civitai.com/models/831971/deep-negative-pony

@Night1176 Woah, thanks a lot for all the info and tips! I will definitely keep coming back to your comment for reference and experimenting! :)

I dont see it either, been enjoying both V3 and V4

why not just using it on tensor.art?

@kuraig maybe because it's not a platform we're familiar with(?, and maybe because we prefer this platform...

First off: thanks for citing my comment! From reading over it seems like you saw some of the things I noticed during the testing.

As for camera angles, I got mixed results as well. But I wasn't trying to use filmic terms, since the model most likely wouldn't recognize that. I might try later this week if I have time.

Old characters I would assume to be hard to prompt without LORAs just for the sheer fact, 90% of older characters are still young looking in Anime, which already has a hard time depicting old. Now remember that most models are fed young girl / guy images. When you see an anime drawing you see 1 of 2 things normally. Near boyish/girlish looking characters, or hello husbando/mommy characters. That's 95% of what's trained, which makes sense. Hence thinking for Elderly, I'd always assume LORA for this.

As for face and body shapes, I have tested this a little just in the last day or two. I have been able to get decent, pregnant belly shots for females, but needs a strong prompt "((Huge Pregnant Belly))" to get a full-term pregnancy in most cases without LORA support. I have gotten some pretty believable chubby / fat characters. I used: "(chubby thighs:1.4), curvy figure" and got some pretty decent results, My issue still was skin tone on V4, I do not like this VAE used. If it's the same one as V3, I really don't know what's adding red into everything. It's still probably my #1 issue with this model over V3... red... everywhere. I was going to try to make a classic fantasy "Dwarf" today [short, stocky, slightly ugly, but kind of attractive], just not sure if I have the time.

As my comment said in what you citied, Unprompted, it turns to mess. If you say for instance: "Pregnant woman, hands on stomach, showing viewer foot, Foot focus, feet focus", you will get the hands 85% of the time with all 5 normal fingers. 65% of the time you will get great detail feet with 6 toes, and then have to work to get 5 toes with equal detail, but this feels like a foot LORA was used, as they aren't plain and basic. When it gets hands and toes right, they are crazy detailed. Or something like waving to viewer, hand comes out clean and clear most of the time. Just to say "Sexy pose / dynamic pose / foot focus / etc..." you will get some janky results, V3 did guess better.


But totally agree, it's good at somethings, bad at others, V4 will have it's specialty. Which to me is: basic anime characters: 10/10 it nails those, Furry: 9/10 nails those without LORAs, basic commands, nails those. Get complex, it comes down to do you have time, or luck. I've seen several amazing images generated just off the Checkpoint below, no LORAs, no additives, just draw model. So it means it can do it, but then look at the prompt, did they get lucky? or was that the 600th gen of that prompt?

@Night1176 I know it's tricky for cps to go for something else than young females as you said. Still I was curious and gave a chance to other pony cps, out of all, "autism" was really faithfull to prompt even at low res.
I even went back to ponyv6 OG and damn was I surprised, the lora support on this bad boy is really great, my images looked so much cleaner. A bit more work needed for composition and backgrounds but way more diverse than "prefect".

What i'm getting from all of this and as i've read around, we should stop overcomplicating checkpoints, well trained loras are way better at handling there specific purpose.
Ponyv6 already has an insane lora support and it just might be the most important thing as of today.
Get some good style loras, use ponyv6 to generate a base image, and use other loras as you need them, and heck you might not even need other loras since pony is already damn strong.

@d970685107384 name one example so I can see.

Name of Character + Show/manga/LN + one simple action and I'll try to generate it. (heads up - WILL NOT do kids. Adult characters only. 1,000 yr old Loli's forbidden too.)

Dude you should try specified Japanese anime character generation, you'd know how bad v4,v3 is.

If it's just letting the AI generate random characters then V4,V3 is fine

@Night1176 You can try a few models from my homepage

@d970685107384 That's not what I asked. I asked you to give me an example of what to generate in V3 or V4 and I'll post my image here of what you say it cannot do. Your page has one image, a girl next to a guys ear.

@Night1176 Try this LoRA

https://civitai.com/models/934143/the-world-is-threatened-by-monsters-sent-by-the-demon

he world is threatened by monsters sent by the Demon

Here's the cue word

score_9,score_8_up,score_7_up，source_anime，best quality, highly detailed, masterpiece, ultra-detailed, absurdres，perfect eyes, perfec face, perfect hands, perfect lighting, 1girl, lodia，Anime-style female mage with long pink hair and side-swept bangs, blue eyes with a flustered, blushing expression. Wearing a white robe with red accents and intricate symbols, holding a golden staff with a dragon ornament and purple gem. Full-body view on a transparent background.

@d970685107384 

https://civitai.com/posts/9423394

For Reference this is official art of the character I was trying to create:

https://www.kaguragames.com/wp-content/uploads/2020/07/Dark-Hero-Party-Lotia.png

I put my full write up in the post.

Just as a heads up: the character's name is Lotia not Lodia, and the game is called Dark Hero Party. This test was destined to fail, but I gave it two attempts which any can see in the post linked above. What you asked for was not possible for any AI to generate with out a well Trained LORA, and yours did not have enough data to do much past that one image. If I downloaded any character LORA with good training, I could have made it any style, any pose, doing actually anything? Goku giving Birth? totally possible, Slime Rimiru being dunked in basketball goal by Kritio, in JoJo's style... would have done it. Why? The LORAs exist.

The issue I feel you have isn't a model issue but content issue. Not all models are made to make everything. What you may be looking to create, might not be possible with Perfect Pony, does it mean it's a bad model no, just not what will work for your needs. Maybe look at other solutions.

You really NEED cell shaded / half-render art to make characters like what was presented with Lotia. Not some ultra high res image with more detail than most artist put into their drawings. Because being in the business, she is flat colored with a little render for shadows.

Also as a FYI: SCORE_9 - SCORE_7 is too high quality for what you are trying to create. You should actually look into what that actually means for Pony. SCORE_9 is some massively detailed image. SCORE_2-4 might where where something like your example is scored. Keep in mind, that was based on the original poster's art preferences. Not a universal adapted score used for all AI prompts. When you get into merged checkpoints, from all I read, you drop all quality tags like that unless in Pony Diffusion. Merged are already highly detailed and far more quality than the base model. So most all of "score_9,score_8_up,score_7_up，source_anime，best quality, highly detailed, masterpiece, ultra-detailed, absurdres" is actually messing with you and not helping. You are making your images far better than what Cell shaded would be rated as.

Anyway! Hope this helps you or any who read this.

o7 to Goofy_Ai once again.

https://civitai.com/posts/9431012

As I commented on another post, Here is just further proof of V4 can make Japanese anime styles. In this case Sword Art Online.

@Night1176 Try this LoRA

https://civitai.com/models/890013/bbq-lover-fallen-priestess-my-sisters-demonic-bloodine-gamecg-style-ponyxl?modelVersionId=995955

@d970685107384 I don't need to try another LORA. I'll get it to look right. Look at the post above your comment. That's V4. Looks like a high quality screen grab from Sword Art Online, minus a few details because I was rushing. I also made it clear, what you were were looking for in the write up on the actual post of images of your LORA. Part of it was you maybe not understanding prompting or trying to create to niche of characters.

Also just a reference, the OP of that LORA you just linked, "autismmix confetti" he used that checkpoint for the images.

"AutismMix_confetti is a small amount of AnimeConfettiTune and AutismMix_pony. Has less style swing than pony and better hands. I prefer this one." - Actually on the Checkpoints page.

"AutismMix_pony is a merge of ponyv6 with loras, its more compatible with certain styles made for the base ponydiffusion model"

So it's merged: Pony Diffusion X Version > Personal Preference LORAS = AutismMix > AnimeConfetti = AutismConfetti. That's a double model merge.

Meaning Perfect Pony XL will NOT handle any image the same way. You can look at the example, I sent you earlier of Lotia. I even used AutismPony as a control test. They look nothing like Perfect Pony. Actually Perfect Pony looks better, a higher quality of image. Not slamming AutismMix, that's a pretty awesome checkpoint. Just for the limited LORA testing, Perfect Pony did better. Maybe add detail LORA and Autism might have looked better. But then it's subjective.

Personally, I believe we've debunked everything you claimed. V2 showed no proof of being "better" for "Japanese Anime Style". With the Screen Cap Style, everything looked closer. Then it comes down to "Artist style". Which "Japanese Anime/Manga/Light Novel/Doujinshi" are you actually looking for. Then it comes down to style. Digital art vs painted art. Older Anime is Painted, modern is Digital. Digital art will always have more details. Painted art will always be more flat. Cell shaded digital is the hybrid concept of that. Trying to make art look painted, but digitally drawn.

Without LORAs of "X" style it's like ordering a burger. Now where is that burger from? Everyone has their version of it. You might get a burger made that's the Whopper from Burger King. But you wanted the Big Mac? Well tell them you want the Big Mac. Get a LORA of the Big Mac and it will LOOK, like that hamburger.

https://civitai.com/images/41379128

It's probably your prompt. Try adding lighting into your prompt somewhere if you feel it's dark. Yes it can be dark in certain scenes from testing. But 75% of the time, that's been user error on my part. Quick prompt change and it's back to normal. Also LORAs can break lighting faster the the raw model.

EX:
"well lit room" "Daylight" "Afternoon" "day time" "bright room"

Neg prompts:

Dark room, dimly lit room, low lighting

@Night1176 thank you very much!🙏🏻

@Ceremon No problem! Good luck!!! Hopefully you can get some good scenes with trying things like this. I will not deny it can make darker scenes than V3, but where needed has been 100% a blessing. Just try to prompt it out. It may not be the model's guessing fault.

Just curious, what did you change?

@PocketPossum not much, just lowering noise multiplier and token merging, switching sampler, adjusting the weight of my lora and playing around with the prompts

@kuraig Thanks for sharing!

any tips moving from 3 to 4 to make it look good? I love 3's style

this model is usually pretty good with hands, but not perfect, negative hand embeddings usually help, but still not perfect https://civitai.com/models/56519?modelVersionId=60938

which version you are using?

@kuraig I'm using the v4 one. If the hand is somewhat complex it goes to shit always. They only appear fine when they are very close and kinda open.

@danilobaz hmm, that's weird. I'm getting good hands almost all the time on v4 with/without lora. Maybe it's your lora, try lowering its weight. Or, you are using low resolution, try 1080x1580

@Goofy_Ai Thanks a lot 👍🏻

@cookiechainsawxx Don't misunderstand him. Goofy only uses that tagger. However, it's still absolutely recommended to train on base models aka ponyV6 instead of any merge model.

@LittleJelly Yip! who are thou stranger? 😱 You seem to know him... "well" as preferences go... Alright mate I'll run one loRA through PerfectPony just to be sure and follow your kind advice, cheers.

@Goofy_Ai Alright, but the samplings I'm seeing with just the first 2 Epochs are something else... This pony is overdue and overfed, sir. Later Goofy 👍🏻

@cookiechainsawxx sounds like your settings are all over the place rather than goofy's checkpoint being bad

@LittleJelly "Something else" as in really good, not bad. Overdue and overfed as a good thing "well trained", don't start fights for no reason, haven't even share my settings, are you felling well?

@cookiechainsawxx its a discussion -_- not a fight... also idk if you are training a lora or a checkpoint and with what scheduler, most ppl moved to CAME+REX, and that one needs at least 6-10 epochs for a decent result when it comes to loras

@LittleJelly The initial question states that I'm working on a loRA, I think sir, than you should avoid jumping into conversations you weren't requested next time and we can all avoid confusion. I don't know where everyone is "moving" as I'm only an amateur and I'm trying to learn but if randoms are gonna take offense out of air I rather use the private message next time, cheers.

@Goofy_Ai And back to you sir, I'm not disrespecting any of your work in any way, I hope that's clear, also I'm not training on 2 Epochs but 30, I was referring to the preview samples OneTrainer gives as the training process goes along. Now is done, gonna test it and share some generations.

@Goofy_Ai I don't understand when did I mentioned any issue at al. I had 0 (zero) problems with this checkpoint or with the training of the loRA. I'm going crazy where did I stated something was wrong, are you English speakers both?. Anyway thanks for the replies.

Huge works better, all pony models were trained with booru tags

@hjaskgdghasda What if you need a long one or a big one? So if I need to get 18cm, 21cm, 25cm? I myself have noticed that Perfect works strangely with sizes. And I use Gelbooru tags.

@Goofy_Ai thx~ღ( ´･ᴗ･` )

yeah its not a lora tho bud

x2