Helios AI Model Powers Real-Time, High-Quality Video
A new AI model named Helios is reportedly generating minute-long, high-quality videos in real-time without requiring heavy optimization. If the claims hold, this represents a significant breakthrough for scalable video generation. The ability to produce content quickly and efficiently could be a game-changer for news platforms needing rapid video turnaround.
Helios, a 14-billion-parameter model, was developed by researchers from Peking University, ByteDance, and Canva. Its code and model weights are publicly available, a significant move in a field with many proprietary models. The model achieves its speed by sidestepping conventional acceleration techniques. It operates without methods like KV-cache, sparse attention, or quantization, which are typically used to speed up large models but can sometimes degrade quality. On a single NVIDIA H100 GPU, the distilled version of Helios can generate video at 19.5 frames per second. This is a substantial leap over other 14B models, some of which drop below 1 FPS or require minutes to render just a few seconds of video. Helios is designed to combat "drifting," a common issue in long-form AI video where consistency in color, position, and objects degrades over time. It achieves this without relying on typical anti-drifting heuristics like keyframe sampling or self-forcing. The system has a unified architecture that supports text-to-video, image-to-video, and video-to-video generation within a single framework. Users can even change the text prompt mid-generation, with the model creating a gradual crossfade between the old and new prompts to avoid a jarring visual break. For evaluation, the researchers developed a new open benchmark called HeliosBench, which contains 240 prompts designed to test real-time, long-duration video generation. In quality scores for long videos, Helios achieved a 6.94, outperforming the previous leading model.
