Helios
PKU-YuanGroup/HeliosReviewed by AI
Checked by RepoPicks Radar on
What is Helios?
Helios is the code and the model weights from a university research paper on long video generation. You give it a text prompt, a picture or a video and it produces a clip that can run for about a minute, at 384x640, in chunks of 33 frames. The authors report 19.5 frames per second on a single H100 card, and a slower low-memory mode that fits in about 6 GB. You can try it in your browser on the authors' Hugging Face demo page, or install it yourself with conda and shell commands on a Linux machine with an NVIDIA card or a Huawei Ascend NPU.
Helios is free and open source (Apache-2.0) and can replace Runway, Kling, and Google Veo 3.
How easy is it to install Helios?
A · download and open · Web
- Linux D
- Web A
The quickest look is the authors' own Hugging Face demo page linked at the top of the README, which needs nothing installed, so the web is an A. Running it yourself is a D: you clone the repository, create a conda environment, pick the PyTorch build that matches your CUDA version, run "bash install.sh", download tens of gigabytes of weights with the Hugging Face or ModelScope command-line tool and start shell scripts. Windows and Mac get no grade: every step is a Linux shell script and the only pointer for an ordinary PC is an unofficial community video.
Helios on GitHub
- Stars
- 2,258on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +11stars in 6 days
Stars over time
Facts
- Runs on
- Linux, Web
- You need to install
- Nothing extra
- To build from source
- Python, Graphics card, FFmpeg
- Disk space
- 1,010 MB–4.5 GB · estimate
- Accounts or keys
- None
- Cost
- Free · The code is Apache-2.0 and the three model versions are free downloads from Hugging Face or ModelScope. Running it fast needs an expensive data-centre card, and the README points at CNAPS.AI, an outside online service that runs Helios for you.
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- Nov 1, 2026
Helios risk signals (6)
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-rrmf-rvhw-rf47, PYSEC-2026-139; transformers: GHSA-fgcw-684q-jj6r, GHSA-x9r9-c232-4q39, GHSA-xrqw-3rrv-vx5w; sentence-transformers: GHSA-jhr6-gm9c-rqjv, PYSEC-2026-4164; accelerate: GHSA-4j2p-28q2-5m79, PYSEC-2026-3804; torch: GHSA-63cw-57p8-fm3p, GHSA-qfhq-4f3w-5fph, GHSA-rrmf-rvhw-rf47
automatic check
- Attentionthe install script removes packages and deletes folders
install.sh runs "pip uninstall triton torchao xformers wandb tensorflow tensorflow-cpu -y" and then "rm -rf ~/.triton/cache/" and "rm -rf /tmp/torchinductor_*". It removes six packages from whatever Python environment is active, including TensorFlow, and deletes folders outside the project, so other AI projects on the same machine can stop working.
AI review
- Infothe speed claim needs a data centre card
README: "minute-scale, high-quality video synthesis at 19.5 FPS on a single H100 GPU (about 10 FPS on a single Ascend NPU)" and "fitting up to four 14B models within 80 GB of GPU memory". The low-memory route says "The Helios model below requires ~6GB of VRAM" through group offloading, and a news line adds that "real-time generation performance depends not only on the GPU, but also on the CPU, memory, CUDA driver version".
AI review
- Infothe models are downloaded separately and are large
The repository has no weights: you fetch Helios-Base, Helios-Mid or Helios-Distilled from Hugging Face or ModelScope with "huggingface-cli download BestWishYSH/Helios-Base". These are 14B-parameter video models, so the real disk use is far above the 1 to 4.5 GB the checks estimate from the repository.
AI review
- Inforesearch code with parts still unfinished
There are no releases, 47 issues are open, the interactive mode warns "⚠️ This feature is still under development — results may not always meet expectations", the Helios-Mid checkpoint "may not meet expected quality", and a news entry says the authors "identified several ways to further improve Helios's performance, such as fixing the i2v train-inference inconsistency".
AI review
- Infoinstall steps pull code straight from git branches
Several routes install unreleased code from main branches: "pip install git+https://github.com/huggingface/diffusers.git", "pip install git+https://github.com/vllm-project/vllm-omni.git" and "pip install git+https://github.com/sgl-project/sglang.git". What you get depends on the day you install it.
AI review
Evidence is shown in its original language.
What Radar could not check about Helios
- How long one minute of video really takes, and how slow the 6 GB low-memory mode is on an ordinary gaming card.
- Real disk space needed for the three model versions plus the Python environment.
- Whether the Hugging Face demo page is up and how long the queue is.
- Whether the install works at all on Windows or a Mac, since no official steps exist for them.
- Picture quality with your own prompts, since the examples and the benchmark are the authors' own.
- What the CNAPS.AI online service charges and what it does with the videos you make there.