The paper addresses the challenge that accelerating video diffusion models is highly instance-specific: the best combination of techniques depends on the model, hardware, and serving configuration. Manually tuning across a large space of options is costly.
Sol Video Inference Engine automates this by using parallel skill agents to optimize each of five acceleration techniques (cross-step cache, sparse attention, token pruning, quantization, and kernel fusion), then an agent integrator composes them into a full stack with human feedback on quality.
The framework was tested on three video models of different sizes (64B, 22B, and 2B parameters) and achieved over 2x end-to-end speedup while maintaining near-lossless VBench quality, demonstrating that an agent-driven approach can significantly reduce human effort in deploying video diffusion models.