Real-Time Long Video Generation (Fastest Variant)
Most AI video tools top out at short clips (5–10 seconds) or slow to a crawl for longer content — and they often need an expensive multi-GPU cluster to keep up. Helios Distilled is the fastest, most efficient member of the Helios family: it turns text prompts, images, or existing videos into fluid, high-quality clips of up to ~60 seconds, generated in real time on a single GPU. Purpose-built for teams that need production-ready video at speed and scale. Built by PKU-YuanGroup.
What it does
Helios Distilled is an AI video generation model that turns text prompts, images, or existing videos into fluid, high-quality video content. It is the fastest and most efficient version in the Helios model family, purpose-built for teams that need production-ready video output at speed and scale. Unlike most competing video AI tools that struggle with longer clips or require expensive multi-GPU infrastructure, Helios Distilled generates up to 60 seconds of coherent, high-quality video in real time.
Problem it solves
- Long-form video at speed – Most AI video tools top out at short clips (5–10 seconds) or slow down dramatically for longer content. Helios Distilled generates up to 60 seconds of video in real time
- Cost-effective infrastructure – Runs on a single GPU rather than expensive multi-GPU server clusters, significantly reducing compute costs
- Flexible creative input – Works from a text description alone, a starting image, or an existing video clip — fitting a wide range of creative and production workflows
- Temporal coherence – Videos remain visually consistent and coherent throughout their full length, without the common "drift" or degradation seen in longer AI-generated clips
- Production-ready deployment – Designed from the ground up to integrate into existing production pipelines and tools
Input/Output
- Input:
- Text to Video: Describe the scene you want in plain language and Helios generates the video
- Image to Video: Provide a starting image and a description of the motion or scene
- Video to Video: Provide an existing video clip and a description to transform or extend it
- Output: A fully generated MP4 video file
- Up to ~60 seconds in length
- High visual quality with smooth, coherent motion throughout
- Parameters:
- Resolution — fixed at 640 × 384 (5:3) — not user-selectable in this build
- Frame count (chunk multiples, default 99 frames (~4s @ 24fps)) — pick from a fixed list of chunk-multiple values that maps to standard clip lengths at 24 fps:
- Output FPS (dropdown, default 24) — choose 16 or 24 fps
24— standard cinematic frame rate16— lighter alternative for shorter or more animation-style output- Advanced Options:
- Random seed (-1 = random, default 1) — set a fixed integer to reproduce the same video across runs
Frames | Duration @ 24 fps |
99 (default) | ~4 s |
198 | ~8 s |
297 | ~12.5 s |
396 | ~16.5 s |
495 | ~21 s |
594 | ~25 s |
693 | ~29 s |
792 | ~33 s |
891 | ~37 s |
990 | ~41 s |
1,089 | ~45 s |
1,188 | ~49.5 s |
1,287 | ~54 s |
1,386 | ~58 s |
1,452 | ~60.5 s |
Frame counts must be chunk multiples — the model generates in 33-frame autoregressive chunks (99 = 3 chunks, 198 = 6 chunks, and so on), which is why the dropdown offers only these specific values.
Accuracy & Speed
Real-time generation speed | 19.5 frames per second on a single H100 GPU |
Maximum video length | ~ 60 seconds |
Infrastructure needed | Single GPU (no multi-GPU cluster required) |
Model Source
- HuggingFace:
Compliance & Provenance
Provider | Open-source (PKU-YuanGroup) |
Provider type | Specialized |
License | |
EU AI Act risk class | Limited Risk |
Art. 50 transparency | Required — outputs are marked. See AI Policy §2. |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.