This is an experimental model !!! | 这是一个实验性模型!!!
UltraVistaNoob 是基于 Noob eps 0.75 的高分辨率模型，优化了超宽屏和大尺寸图像生成。
此模型于2024年9月完成训练
UltraVistaNoob is a high-resolution model based on Noob eps 0.6, optimized for ultra-wide screens and large image generation.
This model was trained in September 2024.
由于训练集画风影响，画师效果会产生偏移，并且在1024x1024分辨率下生成效果会有略微下降。
Due to the influence of the training set's style, the artist's effect may be slightly offset, and the generation effect may be slightly reduced at a resolution of 1024x1024.
Recommend Resolutions: | 本模型推荐分辨率:
1280x1792,1344x1728,1536x1536,1792x1280,1856x1216,1920x1216,1984x1152,2048x1152,2112x1088,2304x1024
这个模型是我突发奇想，想要提升Noob模型的背景表现，所以选择了很多高分辨率背景细节良好的图片。为了尽量保证背景细节，我才用了1536x1536分辨率训练，但其实最后实验发现，背景的细节虽有提升，但是不明显。不过我在训练模型时有了意外收获:
这个模型可以大幅提升在高分辨率和高长宽比分辨率的画面稳定性，包括背景稳定性，和肢体稳定性。虽然仍会有肢体错误的情况存在(我认为可能和Tag也有一些关系),但是频率大幅减少。当人物距离摄像机较远时，面部的崩坏也有所改善(但是过远情况下崩坏仍然存在)。
This model was created out of a sudden urge to improve the background performance of the Noob model, so I selected a lot of high-resolution images with good background details. In order to preserve as much background detail as possible, I trained the model using a resolution of 1536x1536. However, in the end, the experiment found that although the background detail had improved, it was not obvious. But I did get a bonus when training the model:
The model can greatly improve the stability of images with high resolution and high aspect ratio, including background stability and limb stability. Although there are still some limb errors (which I think may also be related to Tag), the frequency has been greatly reduced. The model also improves the face error when the person is far away from the camera (but the error still exists when the person is too far away).