Helios: a real-time 14B video generation model for high-quality minute-scale synthesis

What it solves

Helios addresses the challenge of generating high-quality, long-duration videos (minute-scale) in real-time. It aims to provide a model that is faster and more memory-efficient than smaller models while maintaining strong coherence and quality without needing complex anti-drifting or acceleration techniques.

How it works

Helios is a 14B parameter video generation model that uses an autoregressive approach, generating video in chunks of 33 frames. It employs a multi-scale sampling pipeline (particularly in the Mid and Distilled versions) to increase efficiency. The project includes optimizations for training and inference throughput, supporting features like group offloading to reduce VRAM usage to approximately 6GB and context parallelism (Ulysses, Ring, Unified, and Ulysses Anything Attention) for multi-GPU scaling.

Who it’s for

It is designed for developers and researchers in AI video synthesis who need high-throughput, real-time video generation capabilities, as well as users with consumer-grade hardware who want to generate high-resolution (up to 4K) videos.

Highlights

• Real-Time Performance: Achieves 19.5 FPS on a single H100 GPU (up to 20.89 FPS on optimized hardware).
• Long Video Coherence: Generates minute-scale videos without relying on standard anti-drifting strategies like self-forcing or keyframe sampling.
• Efficiency: Operates without standard acceleration techniques like KV-cache or quantization, yet remains faster than many 1.3B models.
• Versatility: Supports Text-to-Video (T2V), Image-to-Video (I2V), and Video-to-Video (V2V) tasks.
• Broad Compatibility: Integrated with Diffusers, SGLang-Diffusion, vLLM-Omni, and Ascend-NPU.