Add VideoDeltaNet-H3 Model

Model/Pipeline/Scheduler description

VideoDeltaNet (VDN-H3) is a hybrid-attention variant of MiniMax H3 designed for faster video generation with near-lossless quality. It replaces dense video-video attention with a combination of local sliding-window softmax attention and long-range linear attention, while keeping the original H3 backbone largely unchanged.

All weights and training and inference code are available.

Open source status

• The model implementation is available.
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

Weight: https://huggingface.co/OpenVDN/vdn-minimax-h3 Code: https://github.com/OpenVDN/vdn-minimax-h3 I'm the author :)