A team of researchers from Tsinghua University and ByteDance published Helios in March 2026, a video generation model capable of producing 60-second video clips in approximately real time on a single NVIDIA RTX 4090 GPU. The model generates 720p video at 24 frames per second with text-guided control over scene composition, camera movement, and visual style.
Previous video generation models required either multi-GPU clusters or minutes of compute per second of generated video. Helios achieves its speed through architectural innovations that reduce the computational cost of temporal consistency, the process of ensuring that each frame flows smoothly into the next.
Technical Advances Behind Helios
- Generates 720p video at 24 FPS with approximately 1:1 generation-to-playback ratio
- Runs on a single RTX 4090 with 24GB VRAM
- Produces up to 60 seconds of continuous, temporally consistent video
- Uses a novel temporal attention mechanism that scales linearly with video length instead of quadratically
- Supports text-guided control over composition, camera angles, lighting, and visual style
How Helios Achieves Real-Time Speed
Standard diffusion-based video models apply attention across all spatial and temporal dimensions simultaneously. The compute cost grows quadratically with video length, making longer videos exponentially more expensive. Helios introduces a sliding window temporal attention mechanism that processes video in overlapping segments, maintaining consistency at the boundaries while keeping compute cost linear.
Helios achieves real-time video generation by replacing quadratic temporal attention with a linear sliding window mechanism, making 60-second videos practical on consumer hardware.
The model also uses a multi-resolution generation pipeline. It first generates a low-resolution version of the entire video to establish temporal consistency, then upscales in a second pass. This two-stage approach is faster than generating at full resolution from the start because the consistency planning happens at lower computational cost.
Quality Assessment and Limitations
In human evaluation studies, Helios videos were rated as visually coherent 82% of the time for clips under 15 seconds and 71% of the time for clips between 30 and 60 seconds. The quality gap at longer durations reflects the challenge of maintaining visual consistency over extended temporal windows. The most common artifacts in longer videos are gradual color shifts and minor subject deformation.
Helios does not match the quality of larger, slower models like Sora or LTX 2.3 in side-by-side comparisons. The trade-off is explicit: Helios prioritizes speed and accessibility over maximum visual quality.
Research and Creative Applications
The real-time generation capability opens applications that were previously impractical. Interactive storytelling tools can generate video responses in conversation. Design teams can iterate on visual concepts without waiting minutes between attempts. Researchers studying visual perception can generate stimulus videos programmatically at scale.
Helios model weights and code are available on GitHub under an MIT license. Pre-built Docker images for common GPU configurations are also provided.
