No, don't. Quality tagging is there for a reason. Trying to remove it and force "high quality" to be outputted by default destroys many aspects of a model, like creativity and variety, especially variety of style.

Of course, I think pony's quality tagging could be improved, they kind of messed it up.

The quality tagging is fine, but this score aids shit is not compatible with any of the other models and makes it hard to use it as an ingredient. I spent like a month trying to replace it with the known quality tags but the model just breaks, but I think it could be solved with finetuning

@illyaeater Nothing about pony is compatible with other models.

@wewewew It's an sdxl finetune, and other than the furry shit largely uses the same dataset as the other anime finetunes

@illyaeater Less than a quarter of the dataset is the same, probably much less, tagging is different, training method is different. Just check loras trained on one on the other, it just doesn't transfer. The only reason 1.5 anime models were compatible is because they're all based on the nai model.

Anyway, what I mean is, I also prefer the regular best/worst quality tags over the score tags, but some form of quality tagging is necessary. But also, I have no idea why exactly pony diffusion is so much better than other models; who knows if the score system isn't part of the reason, even if it was messed up, so I wouldn't tell them to change it without knowing.

There could be something good with the "_up" tagging method, making it include a wider range of images, instead of say "high quality" being influenced by only images scoring 80 to 100 and "best quality" >100. I assume that instead, for pony diffusion, score_8_up is added to all images scoring 80+, score_7_up 70+, etc.

@wewewew Source for less than quarter?

@illyaeater pony diffusion's description; "roughly 1:1 ratio between anime/cartoon/furry/pony datasets"

Reason I say it's probably much less than a quarter is even in the quarter for anime dataset, pony diffusion and say animagine have very different picks, which you can confirm with some of the lists of common tags, character tags, etc, available. Animagine has mostly sfw or tame nsfw, recent images, mostly vtubers and gacha game characters. Pony has a lot more nsfw, I haven't seen character lists for pony but it looks more balanced. I wouldn't say one is better than other though, they're good at different things. Obviously pony's dataset has resulted in it being better at nsfw, but it's horrible at backgrounds and other things.

@illyaeater @wewewew They probably should have used more natural language at least, several level of "the best quality...", "excellent quality ", "good...", "bad...",, etc.,, which other models could have understood a little bit at least, but 'score_9' is absolutely meaningless for them. I don't think this format has any advantage, except uniqueness, which means that it was more rearely used than 'high quality', but it may probably not matter.

@clex The secret is that it's trained with some pivotal tuning techniques I believe. They've created embeddings which compile multiple tags in one. Score_9 is meant to represent something like 20 individual modifiers that a normal model would use.

And also the reason that it has the issues you identified is largely due to how over-captioned the dataset was. When training, it is most effective to use a balance between captioned and uncaptioned data on the same triggerwords in order to give the model more flexibility. as the model is capable of understanding the differences between a simplified prompt and a complex prompt which are requesting similar style compositions.

If you look into the basics of how ESRGAN models are trained, stable diffusion can be trained in basically the exact same way.

In other words, they could theoretically have added both pony and conventional methods of captioning and combined that with no captions if they had the time and it may have produced a much more well-rounded model.

@illyaeater Instead of trying to replace the score tags with the old quality tags try making something new out of them, lets leave those NovelAI tags with SD1.5. People already made models such as AutismMix and JS2Prony which are based on Pony and are, in my opinion, a lot better than Pony. They did it with score tags as well instead of those quality tags so to me it just sounds like a skill issue honestly.

@GogetaSSGSS3 I mean those are almost just renamed pony models with small differences. The only one that added new shit was 4th tail. Ppl should be actively trying to fix the shortcomings of the model. There is a reason it can't be used with anything else. This score_9 shit is one of the most braindead quality tagging methods I've seen. JS2prony is much nicer looking out of the box, but none of the derivatives improve base pony without losing out on concepts

@Triple_Headed_Monkey I've tried to redirect the score tags into the old tags with leco but any time I got it working with "masterpiece, best quality etc" it broke the model when the tags weren't used. Dunno what else to try atm 

@GogetaSSGSS3 Just to add information, most models like autismmix are very simple merges of models (in this case pony and style loras), so it's not like they "did" much of anything with score tags or anything. Trained models should have the "checkpoint trained" tag instead of "checkpoint merge", although that's set by the uploader with no way to confirm.

@illyaeater Yeah, LECO isn't the way to achieve this. Unfortunately the way to achieve it is through emergent behaviors over successive merging attempts. Through constantly diluting with other models and then adding back in the main model over and over again. Until eventually all of the data is partially corrupted/knocked out of its intended positions and redistributed in a way that allows it all to sit together in the model.

I've been working on it for about 4-5 days, through merging alone. But it looks like Training may be the only option to get it properly moving fast.

But in all honesty, we're just waiting for a decent merge which is capable of using Pony tags on top of conventional prompts. This is a model which will understand more than any other.

@illyaeater I guess we'll have to wait for PonyV7 to come out in a month or 2 cause the score tags are fucked up in V6 compared to the older versions, and this is confirmed by the guy who made it. Not sure if it's true or not but he did say that fixing the score tags is one of the main things he's focusing for this next version. It sucks that we don't know how the score tags work honestly but maybe we'll have more info on it once V7 drops.

@Triple_Headed_Monkey Yeah the distributions being different seems pretty annoying. I haven't tried anything seriously other than a few experimental merges and leco, but the guy that put together some of the merging methods for meh has been trying (unsuccessfully) and kohaku also said maybe he'll give it a go but not sure how that will go since he's always busy with something. Good luck with what you're doing

true, nobody wants to see the low quality score stuff. lol

@GogetaSSGSS3 Creator already stated, skill issue, mistake, PD7 won't do the same mistake. Creator wanted a tag based system score_9 for best visuals, score_6 for medicore, score_3- and below for worst images.

But failed and the result is that you have to put in this long tag line in. Monkey mode was enabled when training was started, and by the time he noticed the mistake, the model was already far too long into training to stop.

@zeal2games549 I know, that's why I'm saying it's not a bad idea, it worked well with the previous Pony versions, it's just that V6 messed it up and now people are getting the wrong idea.

Besides I'm all for people making their own finetunes of Pony instead of trying to combine it with other SDXL models, we already know Pony is already different compared to other models as it is, even if you remove the score tags that wouldn't change the fact that its architecture is different.

@GogetaSSGSS3 The architecture is still SDXL, so merging isn't impossible. It's just annoying as fuck and the retarded tag obfuscation and scuffed score tags certainly aren't helping. But I'm sure sufficiently autistic people will figure it out before the paywalled v7 is released (looking at autism mix who is this four chan guy)

@illyaeater Paywalled v7? Why would it be paywalled? Is this something that was confirmed and I'm unaware of or is it just your assumption? I really don't see why it would be paywalled, most of the sponsors that they have helping them make v7 were because of the massive success of v6 on Civitai and other platforms, I doubt they would make you pay for v7, it just doesn't seem like the right choice, it would be a really shitty thing to do but who knows you might be right

Also yes I know about AutismMix, I really hope the guy who made that can make something that strays away from the classic western base style the model has cause that's my only complaint about Pony thus far

@GogetaSSGSS3 Yeah astralite wrote that v7 might be intentionally gutted (just like v6 with the char/artist tags but maybe worse, who knows) for the public and the paid unlocked version will only be available on discord. In a discord message when he was pissed off, so take that as you will. Wouldn't be all that surprising to try and somehow monetize the success v6 had with people though, but hopefully not. Open source is chugging along but it would suck to lose contributors

I don't think this person responds to comments for some reason. Try messaging them directly

@rqdwdw Thank you for your work! I am going to check the license terms and contact the other model's author as well.

Most likely, I will upload tomorrow or later this week and credit you with a link here.

The merged model I created does have a strong anime style in every generation, but the lineart outlines are sometimes thicker than Pony for Anime, so there is still a use case for both of them. Yours is more capable of a "cute-looking" or "most anime-like" appearance, as with the score_7_up only tag.

@heathergreen95 i'm kind of interested tho, what's the other model that u used to mix with this one? Is it AutismMix? Cause that would have been my choice personally, mainly because this model doesn't deal well with anatomy which is much more consistent with AutismMix. Also when will you be uploading it to the website? I wanna give it a try :)

@GogetaSSGSS3 Hey, I'm so sorry about replying late! I realized only after commenting that the merge might have lost some canon character information due to the new score_ system. I'm still testing models/weights, but I hope to have something uploaded by the end of this week, and I'll let you all know. Thanks for your interest!

This is the other model I talked about in my original comment: https://civitai.com/models/311817