Helios - A real-time long video generation model developed by Peking University in collaboration with ByteDance and other open-source companies.

What is Helios?

Helios is a 14-b parameter real-time long video generation model jointly developed by Peking University and ByteDance, capable of generating minute-level videos at 19.5 FPS on a single H100 GPU. Breakthroughly, the model eliminates the need for conventional acceleration techniques such as key-value caching, quantization, or causal masking. Through innovations like unified history injection, deep compression streaming, and adversarial hierarchical distillation, it reduces computational costs to the level of a 1.3-b model. Helios natively supports text-based video, image-based video, and video continuation, surpassing existing solutions in both quality and speed. The code and model are open-source.

Main functions of Helios

• Text to VideoHelios supports text-to-video generation, directly generating high-quality video content based on text descriptions.
• Image to videoHelios supports image-to-video generation, which can convert static images into dynamic videos.
• video to videoHelios supports video-to-video generation, and can continue or style-transform existing videos.
• Interactive generationHelios supports real-time interactive generation, allowing users to dynamically modify prompts during the generation process.
• Long video generationHelios supports minute-level long video generation and can stably output continuous videos up to 1440 frames per second.

Helios technical principles

• Unified historical injectionHelios transforms a bidirectional pre-trained model into an autoregressive generator by concatenating historical frames with noisy frames as input. The historical frames are kept clean with a fixed time step of 0, while the noisy frames are used for denoising. This means that the same architecture can automatically adapt to three task modes: text-based video when all historical frames are zero, image-based video when only the last frame is non-zero, and video continuation when any historical frame is non-zero.
• Guiding attention mechanismHelios calculates query key values for historical and noisy contexts separately in the self-attention layer, and selectively adjusts the intensity of historical information by amplifying tokens in the head level; in the cross-attention layer, textual semantics are injected only into the noisy context to avoid semantic repetition and superposition of historical frames, and decouple the different statistical characteristics of the two types of contexts.
• Simple anti-drift strategyHelios uses relative position encoding to fix the time index range, eliminating positional offsets in long video generation; it forces the retention of the first frame as a global visual anchor to suppress color drift; and during training, it randomly applies perturbations such as exposure adjustment, noise addition, or blurring to historical frames to simulate error accumulation during inference, eliminating repetitive motion from the source.
• Deeply compressed streamHelios uses multi-stage memory blocks to apply different compression rates to historical frames based on their proximity, keeping the token budget constant. It employs a pyramid unified prediction corrector to sample the latent space in stages at multiple scales, with early low-resolution sampling to determine the global structure and later high-resolution sampling to refine details, reducing the token computation of the 14B model to 1.3B.
• Anti-stratified distillationHelios uses an autoregressive teacher model as a benchmark, obtains multi-scale estimates through phased reverse simulation, achieves course-based learning through dynamic heavy noise scheduling, and introduces adversarial training to break through the performance limit of the teacher model. Finally, it compresses the number of sampling steps from 50 steps to 3 steps without the need for classifier guidance.

Helios project address

• Project official websitehttps://pku-yuangroup.github.io/Helios-Page/
• GitHub repositoryhttps://github.com/PKU-YuanGroup/Helios
• HuggingFace model libraryhttps://huggingface.co/collections/BestWishYsh/helios
• arXiv technical paper: https://arxiv.org/pdf/2603.04379

Applications of Helios

• Film and advertising productionHelios' real-time generation capabilities can significantly accelerate the production process of movie trailers, advertising concepts, and dynamic storyboards, allowing creators to quickly iterate on visual concepts and reduce rendering times from hours to minutes.
• Games and Interactive EntertainmentAs a fundamental component of the world model, Helios supports the real-time generation of game scenes and dynamic environments of unlimited length, providing a coherent flow of visual content for open-world games, virtual reality experiences, and interactive narratives.
• Social media and content creationHelios allows ordinary users to instantly convert text or images into high-quality short videos, lowering the barrier to professional video production and empowering individual creators to quickly produce cinematic visual content.
• Design and PrototypingDesigners can use Helios's image-to-video and video continuation features to transform static design drafts into dynamic demonstrations, preview product animation effects in real time, and accelerate the iteration and feedback cycle.