What is Helios
Helios is a 14B-parameter real-time longvideo generationmodel jointly launched by Peking University, ByteDance, and others. It can generate minute-level videos at 19.5 FPS on a single H100 GPU. The model breakthroughly eliminates the need for conventional acceleration techniques such as KV caching, quantization, or causal masking. Through innovations including unified history injection, deeply compressed flow, and adversarial hierarchical distillation, it reduces computational cost to the level of a 1.3B model. Helios natively supports text-to-video, image-to-video, and video continuation, surpassing existing solutions in both quality and speed. The code and model have been open-sourced.
Main features of Helios
-
Text-to-video: Helios supports text-to-video generation, directly generating high-quality video content based on text descriptions.
-
Image-to-video: Helios supports image-to-video generation, which can transform static images into dynamic videos.
-
Video-to-video: Helios supports video-to-video generation, enabling continuation or style transfer for existing videos.
-
Interactive generation: Helios supports real-time interactive generation, allowing users to dynamically modify prompts during the generation process.
-
Long video generation: Helios supports minute-level long video generation, stably outputting coherent videos of up to 1440 frames.
Technical principles of Helios
- Unified history injection: Helios concatenates historical frames with noise frames as input, transforming a bidirectional pretrained model into an autoregressive generator. The timestep of historical frames is fixed at 0 to keep them clean, while noise frames participate in denoising. A representation control mechanism enables the same architecture to automatically adapt to three task modes—when all history is zero it is text-to-video, when only the last frame is non-zero it is image-to-video, and when any historical frame is non-zero it is video continuation.
- Guided attention mechanism: Helios separately computes query-key values for historical and noise contexts in the self-attention layers, selectively regulating the strength of historical information through head-level amplification tokens. In the cross-attention layers, text semantics are injected only into the noise context, avoiding repeated semantic superposition of historical frames and decoupling the different statistical properties of the two types of context.
- Simple anti-drift strategy: Helios adopts relative position encoding to fix the temporal index range, eliminating positional drift in long video generation. It forcibly retains the first frame as a global visual anchor to suppress color drift. During training, perturbations such as exposure adjustment, noise addition, or blurring are randomly applied to historical frames to simulate error accumulation during inference, eliminating repetitive motion at the source.
- Deeply compressed flow: Helios uses multi-stage memory chunking to apply different compression rates to historical frames according to their temporal distance, keeping the token budget constant. It employs a pyramid unified prediction corrector to sample in stages across multi-scale latent spaces, determining global structure at low resolution in the early stage and refining details at high resolution in the later stage, reducing the token computation of the 14B model to the level of 1.3B.
- Adversarial hierarchical distillation: Helios uses an autoregressive teacher model as the baseline, obtains multi-scale estimates through staged reverse simulation, implements curriculum learning via dynamic renoising scheduling, and introduces adversarial post-training to break through the performance ceiling of the teacher model, ultimately compressing the number of sampling steps from 50 to 3 without the need for classifier guidance.
Project address of Helios
- Project website：https://pku-yuangroup.github.io/Helios-Page/
- GitHub repository：https://github.com/PKU-YuanGroup/Helios
- HuggingFace model library：https://huggingface.co/collections/BestWishYsh/helios
- arXiv technical paper：https://arxiv.org/pdf/2603.04379
Application scenarios of Helios
- Film and television and advertising production: Helios's real-time generation capability can greatly accelerate the production workflows of movie trailers, advertising creative concepts, and dynamic storyboards. Creators can rapidly iterate visual concepts, shortening traditional rendering times of several hours to the minute level.
- Games and Interactive Entertainment: As a foundational component of world models, Helios supports the real-time generation of game scenes and dynamic environments of unlimited length, providing a coherent stream of visual content for open-world games, virtual reality experiences, and interactive storytelling.
- Social Media and Content Creation: Ordinary users can use Helios to instantly transform text or images into high-quality short videos, lowering the barrier to professional video production and empowering individual creators to quickly produce cinematic-level visual content.
- Design and Prototype Validation: Designers can use Helios's image-to-video and video continuation features to transform static design drafts into dynamic presentations, preview product animation effects in real time, and accelerate the iterative feedback loop.