Pony scoring system concerns.
There is a problem in using flawed aesthetic scoring system introduced in Pony models. First of all, infamous "Pony AI slop" look introduced with multilayered tags like score_9, score_8. Using 10 overlayed quality tags is not a good idea. Contaminating captions with "subjective" aesthetic preferences means so-called good anatomy and composition will be locked behind those high-score score_9 tags. But with those comes the slop element. These tags overpower artist tag, destroy the intended shading. And even at this preview stage it's bleeding influence is visible. Sure, you can mitigate stylistic influence using something like score_6, but then anatomy and/or composition will suffer instead. There is a reason NovelAI uses simple quality descriptors. Along 2d anime waifu generating pony score system is both bane and laughing stock for these reasons. And to be completely fair, creator of Pony did not understand what he was doing with this scoring, messing up the captions by labelling every image with long "score_9, score_8, score_7... tags" sequence. Overall it's your decision to use them or not.
Sorry for ranting.
I hate the score system 😭😭😭😭 Since I’ve never used Pony models, I was pretty confused for a while about what score actually was. The main reason I stay away from Pony is exactly because of that typical AI slop it has , please don't use score 😭😭😭
The quality tags are randomly dropped out during training at 50% probability. The model trains on images with no pony tags as frequently as it trains on images with them. As such they are entirely optional, it's just another dimension available to you to control aesthetics, if you want to use them.
can you also clarify the time ranges of the newest, recent, etc tags?
The quality tags are randomly dropped out during training at 50% probability. The model trains on images with no pony tags as frequently as it trains on images with them. As such they are entirely optional, it's just another dimension available to you to control aesthetics, if you want to use them.
Is this also true for low quality tags?
The quality tags are randomly dropped out during training at 50% probability. The model trains on images with no pony tags as frequently as it trains on images with them. As such they are entirely optional, it's just another dimension available to you to control aesthetics, if you want to use them.
To be honest I am not quite convinced. You are training a rather small model, thus learnable concepts are limited by design (which is okay in general for such a focused model). I see the following drawbacks of the pony tags in addition to the ones already mentioned by @shiboishi :
- they occupy capacity for more usable concepts
- they are not documented on danbooru, thus we do not clearly know how they will affect generations and are left with trial and error
- they will slow down model convergence in training as they need to be learned
- they might adversly affect niche tags which are not present that frequent in the dataset, e.g. tags only appearing on 50-100 images, in case the statistics demon hits and dropout normal distribution is shifted for the tag (can be mitigated by appropriate means probably)
I understand people can throw prompts from their old pony generations at Anima, so that's a plus. But still I'd rather deprecate the score tags and only label images with "real" concepts.
since there is more example on civita now. I checked and tried most of those it seems score tags does impact the image quality a lot.
seems it does mess up prompts usage somehow.