ByteDance, in collaboration with researchers from Peking University and Canva, officially released Helios, a groundbreaking 14-billion parameter video generation model.
The release is particularly significant because it is open-source under the Apache 2.0 license, allowing for commercial use, modification, and redistribution. Helios is designed to solve the two biggest “pain points” in AI video: the slow speed of generation and the “drifting” (visual degradation) that usually occurs in videos longer than 10 seconds.
The “Real-Time” Breakthrough
Helios is the first 14B model capable of “Real Real-Time” generation. While most high-end models take minutes to produce a few seconds of footage, Helios can generate video faster than you can watch it.
- Inference Speed: It runs at 19.5 FPS on a single NVIDIA H100 GPU.
- Distillation Magic: By using “Adversarial Hierarchical Distillation,” the model collapses the traditional 50-step diffusion process down to just 3 steps without losing significant visual quality.
- Efficiency: Remarkably, it achieves this speed without using standard industry “shortcuts” like KV-caching, quantization, or sparse attention.
Solving the “Minute-Scale” Challenge
Most AI video models struggle with “drift”—where a person might suddenly grow a third arm or the background melts after 5 seconds. Helios is built to maintain coherence for up to 60 seconds of continuous footage.
| Feature | How Helios Solves It |
| Long-Video Drift | Uses Relative RoPE and First-Frame Anchoring to keep the subject consistent for 1,400+ frames. |
| Memory Usage | Infrastructure optimizations allow it to fit four 14B models within a single 80GB GPU. |
| Unified Tasks | A single architecture natively supports Text-to-Video (T2V), Image-to-Video (I2V), and Video-to-Video (V2V). |
Available Model Variants
ByteDance has released three distinct versions of the model on Hugging Face to suit different hardware and quality needs:
- Helios-Base: The “gold standard” for quality; uses 50 steps for maximum fidelity.
- Helios-Mid: Uses a “Predictor Corrector” method to balance speed and quality.
- Helios-Distilled: The ultra-fast version that hits the 19.5 FPS mark using only 3 steps.
Where to Find It
The project is fully accessible to the developer community:
- Weights: Available on Hugging Face (
multimodalart/Helios-Distilled). - Code: Hosted on GitHub (
PKU-YuanGroup/Helios). - Integration: Features “Day 0” support for ComfyUI, Diffusers, and vLLM-Omni.
