Yuan Wentao

Thanks for your great work! In supplementary material, instead of Post-LN layer, you propose an encoder layer variant which has fast convergence without extensive hyperparameter tuning. The difference of your...

Redundant homography matrix inversion

1

Thanks for your great work! In the original paper of [Unsupervised deep homography](https://arxiv.org/pdf/1709.03966), authors use the spatial transformation layer to perform inverse warping to get the warped image. If I'm...

## Detailed description I am confused about the implementation of absolute depth calculation in [omnidir.cpp](https://github.com/opencv/opencv_contrib/blob/8dfeed73b9389ba09411498dc1e923431d7fe292/modules/ccalib/src/omnidir.cpp#L1489), which is different from the original paper (**Binocular Spherical Stereo**). specifically, the depth is calculated...

Thanks for your great contribution ! I am confused about your implementation of absolute depth calculation, which is different from the original paper. In [Line 1483](https://github.com/jiuerbujie/omnidirectional_calibration/blob/3ac96ed8d4ba3d6eb069e35e0ee2f30d1923bc53/src/omnidir.cpp#L1483) of omnidir.cpp, the depth...

Temporal MultiDiffusion doesn't work

8

Thanks for your great job! I implemented the Temporal MultiDiffusion based on the paper, but there are still significant appearance differences between different video clips. The result shown in Figure...

torch 2.5 CuDNN backend for SDPA NaN error

2

### Describe the bug When using the recently released PyTorch 2.5, the default SDPA backend is CUDNN_ATTENTION. In the example's CogVideoX-lora [training script](https://github.com/huggingface/diffusers/tree/main/examples/cogvideo), NaN gradients occur right at the first...