PKU-YuanGroup/Helios

Helios: Real Real-Time Long Video Generation Model

What it solves

Helios addresses the challenge of generating high-quality, minute-scale long videos in real-time. It eliminates the need for complex anti-drifting strategies (which prevent visual inconsistencies over time) and standard acceleration techniques while maintaining strong coherence and high inference speeds.

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 pyramid-based sampling pipeline and specific optimizations to improve training and inference throughput. The project provides three model versions: Base (highest quality), Mid (intermediate), and Distilled (highest efficiency). It also supports advanced memory and compute optimizations like group offloading to reduce VRAM usage to ~6GB and context parallelism (e.g., Ulysses, Ring, and Unified Attention) for multi-GPU scaling.

Who it’s for

It is designed for AI researchers and developers working on video synthesis who need a balance between high-quality long-form video generation and real-time performance on available hardware (from consumer-grade PCs to H100 GPUs).

Highlights

  • Real-Time Performance: Achieves up to 19.5 FPS on a single H100 GPU.
  • Long Video Synthesis: Capable of generating minute-scale videos without relying on conventional anti-drifting methods.
  • Memory Efficiency: Supports group offloading to fit the model in as little as 6GB of VRAM.
  • Versatile Modalities: Supports text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V).
  • Broad Integration: Day-0 support for Diffusers, SGLang-Diffusion, vLLM-Omni, and Ascend-NPU.

Related

  • Project
    Tencent-Hunyuan/HunyuanVideo-1.5

    A lightweight 8.3B parameter video generation model that enables high-quality text-to-video and image-to-video synthesis on consumer-grade GPUs.

  • Project
    SandAI-org/MAGI-1

    MAGI-1 is an autoregressive video generation model that produces high-fidelity videos chunk-by-chunk to ensure temporal consistency and physical accuracy.

  • Dispatch
    State of Open Video Generation Models in Diffusers

    Hugging Face provides a comprehensive overview of open video generation models and introduces a suite of Diffusers optimizations that can reduce VRAM requirements for models like HunyuanVideo from 60GB to approximately 6.5GB.

  • Project
    kairos-agi/kairos

    Kairos is a 4B-parameter cross-embodiment world model that unifies video generation and executable action prediction for real-time robotic control across diverse hardware platforms.

  • Project
    ModelTC/LightX2V

    LightX2V is an open‑source inference framework for fast image and video generation (T2V, I2V, T2I, I2I, T2AV, etc.). It supports many state‑of‑the‑art models (MiniMax‑H3, Wan, HunyuanVideo, Qwen‑Image, etc.) and provides speed‑up techniques such as 4‑step distilled LoRAs, FP8/NVFP4 quantization, tensor/sequence parallelism, and disaggregated deployment on a wide range of hardware. Benchmarks show up to 3.9× faster inference than competing frameworks on H100 GPUs. The project offers Docker images, pip install, Gradio/ComfyUI front‑ends, extensive docs, and an online demo.