Video generation models are entering a new battlefield: real-time long video. Peking University, ByteDance, Canva, and others have jointly released Helios, a 14-billion-parameter unified video generation model capable of real-time, minute-long video generation at 19.5 FPS on a single H100 GPU, without requiring standard acceleration techniques like KV-cache, sparse attention, or quantization.
Where is the core breakthrough?
Traditional video generation models face a drifting problem in long-video scenarios — by mid-to-late generation, image quality seriously degrades. The industry usually mitigates this through self-forcing, error-banks, or key-frame sampling, but these methods introduce extra compute overhead. Helios' solution is from the training side: the team analyzed typical failure modes, explicitly simulating drifting scenarios in training, letting the model learn to eliminate repetitive motion from the source, rather than post-hoc remedies.
Another key innovation is efficiency. Unlike small models, 14B-parameter models previously couldn't run real-time on a single card. Helios reduces compute cost to be on par with or lower than 1.3B-parameter models by recompressing history and noise context, reducing sampling steps, while infrastructure-level optimization further reduces VRAM usage — 4 14B models can fit in 80GB VRAM for parallel training.
How does actual performance stack up?
Across both short video (121 frames) and long video (1440 frames) dimensions, Helios surpasses previous distilled models, with inference speed significantly leading same-scale solutions. It supports T2V (text-to-video), I2V (image-to-video), and V2V (video-to-video) tasks, unified-modeled within a single architecture.
The team announced it will open-source code, base models, and distilled models. This means the community can reproduce and optimize this capability on consumer-grade hardware, with profound implications for the popularization of video generation technology.
Helios' emergence validates an important trend: large-model inference doesn't have to use only the compromise path of small model + distillation. Through joint design of training strategy and system optimization, 14B-class models can also achieve real-time interaction at controllable cost. This has structural impact on multimodal Agents, video editing tools, and even real-time video dialogue scenarios. Original: https://arxiv.org/abs/2603.04379 | https://pku-yuangroup.github.io/Helios-Page