ByteDance open-sources Helios for real-time video

Helios is an open-source 14B real-time long video generation model from ByteDance. Built for high throughput, it hits 19.5 FPS on a single Nvidia H100 GPU, churning out coherent video clips up to one minute long without drifting into visual chaos.

What it can do

Helios natively handles text-to-video, image-to-video, and video-to-video tasks.

The headlining metric is its raw generation speed. In terminal tests, a clip spanning over 30 frames renders in under two seconds. Its distilled variant clocks in around 19.5 frames per second, reaching the threshold of practical real-time delivery.

Durability over time is the other major win. Most real-time models quickly disintegrate into blurry artifacts or repetitive loops when stretched beyond a few seconds. Helios maintains coherent stylistic stability for up to a full minute.

Because it prioritizes real-time performance, you will still spot slight noise and softer edges compared to slow, heavyweight models. However, it noticeably outclasses existing real-time options and stands shoulder to shoulder with several non-real-time baselines.

How it works and requirements

Under the hood, Helios runs on a 14B autoregressive diffusion architecture.

Instead of piling on standard acceleration techniques like KV-cache, sparse attention, or heavy quantization, the team kept the architecture remarkably lean. They heavily compressed historical and noisy contexts while slashing sampling steps, bringing computational costs down to the ballpark of a 1.3B model.

To cure video drifting, the researchers avoided external heuristics like error banks or keyframe sampling. Instead, they designed training routines that explicitly simulate drifting, neutralizing repetitive motions directly at the source.

The hardware requirement is the only steep catch: hitting that 19.5 FPS speed requires an Nvidia H100 GPU. That said, its memory management is exceptionally light, allowing training batches within 80 GB of VRAM without complex sharding frameworks.

The project is fully open-sourced on GitHub, complete with execution pipelines, local setup guides, and training code, with base and distilled model checkpoints lined up for community use.

Why it matters

If you have access to enterprise hardware, Helios is definitely worth testing right away. It proves that a 14B model can pull off minute-long, real-time generation on a single card without relying on cumbersome engineering workarounds.

We think this shifts AI video closer to responsive, interactive generation. Moving from long rendering waits to near-20-FPS delivery lays the groundwork for real-time digital avatars and interactive virtual environments down the road.