my test case for comfyui workflow is bad

[zhouhualei]

here is my test image and video：

https://github.com/fudan-generative-vision/champ/assets/968848/86637a99-5e30-45cc-9052-45f7971c53d0

and here is the result video：

https://github.com/fudan-generative-vision/champ/assets/968848/25df2f66-e1eb-4488-82cf-7c8258349f7a

So, what's the problem? is there any requirements for test data? Thanks

[Leoooo333]

Hi! It's a great ComfyUI Wrapper for Champ. But for the Semantic_map, we actually use a different ColorMap and different part segmentation. This "gap" may introduce severe artifacts for diffusion models. And I guess for this example, the normal map and pose predicted are severe, would you like to show them here?

[zhouhualei]

@Leoooo333 ok, the followings are the motion sequences:

[Leoooo333]

Hi there, thanks for your detailed description. It's a great ComfyUI Wrapper for Champ. But for the Semantic_map, we actually use a different ColorMap and different part segmentation. This "gap" may introduce severe artifacts for diffusion models. We are going to release our SMPL & Rendering code, which can solve this gap.

[subazinga]

How did you achieve these condition maps? I think there may be a large domain gap between these maps with ours (which are achieved by SMPL).

[zhouhualei]

@suqingkun 要不直接用中文沟通？我是根据你们的视频教程，跑出来的效果，所以不太清楚问题在哪里？https://www.youtube.com/watch?v=cbElsTBv2-A

我这边完整工作流截图是这样的，还需要我提供哪些信息吗？

[Leoooo333]

@zhouhualei hello 你好，这个comfyui有在README上面提到是unofficial的，所以是存在一点问题的，主要是semantic map那一块的color map以及part segmentation不同。你的例子中的motion人物小，动作太快，我看到产出的几个condition中在时间上不连续，伪影比较严重，这些是受到normal，depth预测网络能力的约束，而不是Champ的问题。建议您使用我们新发布的SMPL & Rendering里的教程去准备数据。

[zhouhualei]

@Leoooo333 预处理的代码和教程已经发布了吗？看你们roadmap里好像delay了

[Leoooo333]

之前发布之后对国内配置环境不太友好，就revert了。现在的这个版本正在PR中，你也可以直接fork这个branch：https://github.com/Leoooo333/champ/tree/feature/data_processors

[zhouhualei]

之前发布之后对国内配置环境不太友好，就revert了。现在的这个版本正在PR中，你也可以直接fork这个branch：https://github.com/Leoooo333/champ/tree/feature/data_processors

好的，我试试，感谢！