Lights, Camera, Inference! Video Generation as a Service With VLLM-O... Ricardo Noriega & Doug Smith

This talk covers the exploration of video generation as a service using vLLM-Omni and the LTX-2 open weights video model. The speakers, Ricardo Noriega and Doug Smith from Red Hat, discuss their experimental stack designed to understand the potential of a multimodal approach for production use. They delve into the high-performance autoregressive inference capabilities of vLLM and its extension to support LTX-2, detailing the integration with front ends and scalable deployment processes. Additionally, the presentation addresses the transition from novelty content to practical applications, such as creating consistent characters and interactive storytelling, while identifying the challenges that remain for making generative video production-ready.