此模型建议搭配画风lora出图，虽然目前只有我做的sakimichan版测试模型比较确定效果。科普：xl大模型之间有层可悲的lora不通用的壁障，同一个大模型微调来的，壁障薄一点。由7成青叶大佬的rev3和3成anime31融合而成。优点：根据我的推测应该是一个比较容易优化，且色图出图成品率还不错的一个大模型。缺点：画风和动作仍然需要一定的优化，要是控制能更简单又不失去这种泛化性就好了。不过还是不确定，毕竟是刚测试出来的大模型，改天我试试能不能优化它。作者为什么要弄一个这个模型出来。以下皆为推测，作者还在验证中，概不负责。个人推测，做个比喻，普通的大模型就像一块橡皮泥，一般的优化是给橡皮泥捏出了雏形。pony大模型就像一个成品的积木，不知道内部的结构，不能随便往里面加东西，否则积木会塌。在使用pony的过程中，我发现它稳定的同时有点僵硬，它就像一个内置了近百个动作模型和上百个角色模型的大模型。但是我起初觉得这并没有问题。但是听说有大佬试图优化pony的时候容易崩肢体，那大概理解了，有可能就是新加入的素材数据把内置的动作模型冲散了，通过固定动作稳定的肢体自然也就不稳定了，不过这些都是推测。基于这个推测。对于pony制作组来说继续优化肯定不成问题，因为一个模型稳定十个动作是僵硬，稳定100个动作是不错，稳定1000个动作是灵活，稳定10000个动作是几乎人体万能。但不知道pony素材是按什么比重结构的人去优化它，则可能更容易搞坏它。然后我就知道了琥珀青叶大佬的kohaku rev3模型，当天试了一下，很难控制，但是肢体整体不坏，同样是稳定到了手指数量的固定还有些困难的地步，但是容易出一些构图有点抽象的图，整体没有问题，但还需要一些微调来增加使用的方便性和良品率，我想到了把pony模型当成一个超级色图画师模型少量加进去当微调。但是，加进去之后直接是大片的抽象和一个颜色的图块，简直离谱，试了两个都是这样的，于是加入了三成的anime31，舒服多了。但这个模型给人的感觉还是更接近一个胚胎。不过估计也没什么人会注意到这个模型。我拿来自娱自乐一下吧。This model is recommended to be designed with the Lora art style, although currently only my Sakimichan version of the test model is more certain of its effectiveness.Science popularization: There is a layer of sadly incompatible barriers between XL large models, and the barriers are thinner when fine tuned from the same large model.Formed by the fusion of Rev3, which accounts for 70% of Aoba experts, and Anima31, which accounts for 30%.Advantages: Based on my speculation, it should be a large model that is relatively easy to optimize and has a good yield rate in color mapping.Disadvantage: The art style and actions still need some optimization. It would be great if the control could be simpler without losing this generalization ability.But I'm still not sure, after all, it's a newly tested large model. I'll try optimizing it another day.Why did the author create this model.The following are speculations, and the author is still verifying them and is not responsible.My personal speculation is that, as a metaphor, an ordinary large model is like a piece of rubber clay, and the general optimization is to create a prototype for the rubber clay.The Pony model is like a finished building block, without knowing its internal structure, you cannot add anything to it casually, otherwise the building block will collapse.In the process of using Pony, I found that it is stable but a bit stiff, like a large model with nearly a hundred action models and hundreds of character models built in. But at first, I didn't think it was a problem. But I heard that some experts are prone to breaking limbs when trying to optimize Pony. That's probably understandable. It's possible that the newly added material data has scattered the built-in action model, and the limbs stabilized by fixed actions are naturally unstable. However, these are all speculations.Based on this speculation.For the Pony production team, continuing to optimize is definitely not a problem, because stabilizing ten actions for a model is stiff, stabilizing 100 actions is good, stabilizing 1000 actions is flexible, and stabilizing 10000 actions is almost universal for the human body. But those who do not know the proportion structure of Pony material to optimize it may be more likely to damage it.Then I learned about the Kohaku rev3 model of Amber Leaf Master. I tried it out that day and it was difficult to control, but the overall body condition was not bad. It was also difficult to stabilize to a fixed number of fingers, but it was easy to produce some abstract compositions. Overall, there was no problem, but some fine-tuning was needed to increase usability and yield. I thought of adding a small amount of Pony model as a super color artist model for fine-tuning. However, after adding it, there were immediately large abstract blocks and a single color block, which was simply outrageous. I tried both and it was like this, so I added 30% anima31, which was much more comfortable. But this model still gives a feeling closer to an embryo.However, it is estimated that few people will notice this model. Let me entertain myself with it.

Files with SHA256: 1031cd901cbd77887f1bde10c56c5a81f7d494a3885aadb64f2206e33049c1df