Phr00t

/

Qwen-Image-Edit-Rapid-AIO

like 1.92k

When using V13 for image generated videos, the facial features of the character at the beginning of the video can maintain good consistency with the facial features of the character in the initial image, but quickly become different facial features and subsequently become fixed. This is very serious and directly leads to a poor user experience for character generated videos, which cannot be used normally. And the remix high-low noise model next door can always maintain consistency. But the remix model is too complex, and I genuinely want to use a simple model like Rapid AIO. I don't know if the latest V14 and V14.1 versions have completely solved the consistency problem of character appearance and body part features? In addition, it is hoped that repid aio can optimize the degree of adherence to prompt words, such as "the character turns 360 degrees in place and eventually returns to the front facing the camera, and during the turn, the head and limbs turn in linkage". In such prompt words, it is necessary to add the sentence "during the turn, the head and limbs turn in linkage", otherwise the limbs will twist and become confused. Even if this prompt is added, it will be found that after the character turns 180 degrees away from the camera, their face will abruptly turn back to the front, which is very frightening. I hope Repid aio can optimize in this regard.

Summary:

1. There is a problem with the consistency of appearance of the characters in the TuSheng video
2. The model's adherence or understanding of prompt words is insufficient.
   It is recommended to conduct detailed testing when integrating the new Loras model with Rapid AIO, as many people are using this model and looking forward to it

使用V13进行图生视频时，视频开始人物脸部与起始图像里的人物容貌特征能保持很好的一致性，但是迅速会变成另外的容貌特征并后续固定。这非常严重，直接导致人物图生视频的体验很差，无法正常使用。而隔壁的remix高低噪模型就能始终保持一致性。但是remix模型太复杂，我是真心想使用rapid-aio这种简单模型。不知道最新的V14和V14.1版本是否彻底解决角色容貌、身体部位特征一致性问题？另外提示词的遵循程度，希望repid-aio能优化，比如“角色原地转身360度，最终回到正面朝向镜头，转身期间头与四肢联动转向”，这种提示词里，必须要添加“转身期间头与四肢联动转向”这句话，否则四肢会扭曲错乱。即使添加了这一句提示词，会发现角色转身180度背朝镜头后，脸部会很突兀地直接变回正面，非常吓人。希望repid-aio在这方面能优化。

总结：
1、图生视频角色容貌一致性存在问题
2、模型对于提示词的遵循或理解不足。
建议rapid-aio整合新loras模型时，最好进行细致的测试，毕竟很多人在用这个模型，很期待