FastVideo Opens Dreamverse: Real-Time Video Generation Framework Now Available

Summary

FastVideo has open-sourced Dreamverse, its real-time video generation workspace, releasing the frontend and backend as a reference application for developers building generative video systems. The platform enables "vibe directing"—an iterative workflow where creators use natural language to steer video generation, keeping creators in the creative loop rather than sidelining them during generation.

Dreamverse is built on LTX-2, an open-weights diffusion model, and is optimized to run on a single NVIDIA B200 GPU, making it accessible for developers who want to build their own real-time video applications. The open-source release includes the browser workspace, FastVideo backend runtime, NVIDIA Blackwell-optimized generation paths with NVFP4 inference and FA4, prompt rewriting capabilities, tests, benchmarks, and Docker images for both development and production deployment.

The platform supports multiple deployment scenarios—local GPU, self-hosted B200 servers over SSH, Docker containers, or serverless Modal—making it flexible for different use cases. FastVideo also provides a mock backend for frontend development without hardware, allowing developers to prototype workflows before access to expensive GPU resources. This release positions Dreamverse as both a runnable product prototype and a reference architecture for the broader community building real-time video generation systems.

• Self-hostable architecture supports local deployment, cloud GPU providers, Docker, or serverless platforms; includes mock backend for development
• Reference architecture enables developers to build their own real-time video generation and editing applications within FastVideo ecosystem