FastVideo launches Dreamverse, its real-time video "vibe directing" tool

UCSD's Hao AI Lab released Dreamverse, a real-time video vibe directing tool for FastVideo using LTX-2 and NVIDIA B200 GPUs to stream interactive renders.

UCSD's Hao AI Lab open-sources Dreamverse, its real-time video generation environment developed as part of the FastVideo project. Both the frontend and backend are released as a reference application for anyone wishing to build their own generative video systems.

The tool leverages 'vibe directing', a video adaptation of the vibe coding logic: one starts with a simple idea, observes the result, then guides the next steps using natural language. Maintaining the subject, modifying the camera, extending the scene, or branching in another direction is done within a short iteration loop, with the rendering displayed as generation progresses rather than segment by segment.

Dreamverse is based on LTX-2, LTX's diffusion model, and runs on a single NVIDIA B200 GPU, with each session utilizing one card. The provided stack brings together the browser workspace, the runtime that manages sessions, prompt memory, security, and rewriting, a pool of GPU workers, and an fMP4 streaming layer transmitted via websocket. A security filter relies on fastText classifiers to detect explicit content and hate speech, while a rewriting module enriches each instruction with camera, movement, and context details via low-latency LLM endpoints such as GroqCloud.

Regarding optimization, generation leverages torch.compile, FA4 attention designed for the Blackwell architecture, and NVIDIA's NVFP4 format to better utilize the B200's Tensor Cores. Each segment retains the last frames and audio as conditioning, which ensures visual and auditory continuity from one shot to the next. The team is now exploring methods to run Dreamverse on consumer-grade cards like the RTX 5090, 4090, and 3090.