VDN-H3: Hybrid Attention Model MiniMax H3 Generates Video Faster Than Real-Time (Note: H3 Model Unavailable in South Korea)

The VDN-H3 research team significantly improved video generation speed by applying a hybrid attention architecture to the MiniMax H3 backbone. To resolve the bottleneck where traditional Softmax Attention accounted for over 85% of execution time, they adopted a method that processes proximate frames with Softmax Attention and summarizes distant context with bidirectional Linear Attention.

Performance Improvements and Benchmarks Measured at 768p resolution for a 14.4-second video generation (denoising phase), the model achieved 74.5x faster performance compared to Dense H3 (1 GPU).

• 1 GPU: 16.74s per NFE → 6.41s (with FP8 applied)
• 8 GPU Distributed: 1.40s per NFE (generation faster than playback time)
• Block-level Acceleration: 332.5ms → 125.3ms on B200 (2.65x acceleration)

Core Technology and Restrictions Video Delta Attention (VDA) is presented as the core contribution, ensuring the stability of state updates through frame-wise normal equations. Training involved three-stage alignment and DMD2-based distillation. However, the weights follow the MiniMax H3 Community License, which prohibits use, reproduction, and distribution in South Korea, the United States, the EU, and the United Kingdom. Users in South Korea can only refer to the code, and using the actual model weights requires separate negotiation.