docs · getting started

From zero to your first film

OpenVideo v0.1.0 runs local MiniMax H3 via ComfyUI on your GPU — install, pull, run, plus an agent skill for high-quality prompts. No cloud account. The VLM judge is opt-in (honest SKIPPED when unset); the multi-shot director is experimental, not finished product.

What you need: Linux or macOS, an NVIDIA GPU (~10 GB+ VRAM practical floor; 21 GB+ comfortable for INT8), ~60 GB free disk for weights, Python 3.10+, ffmpeg, and a good network for the first download. Without a GPU you can still explore the UI mockup — real generation needs a local backend.

1. One-command install

This is the fastest path. It sets up ComfyUI, downloads H3 weights (resumable), and can run a first test clip.

curl -fsSL https://open-video.ai/install | bash

Windows (PowerShell):

irm https://open-video.ai/install.ps1 | iex

Installer auto-picks H3 quant from VRAM (nf4 / w4 / int8). Prefer flags: bash install.sh --self-test · bash install.sh --help

OS	Path	Notes
Linux	`curl …/install \| bash`	Full H3 + ComfyUI
macOS	same curl	Setup + dry-run; gen is community/MLX
Windows	`irm …/install.ps1 \| iex`	WSL2 for GPU; native dry-run

1b. Other install paths

OpenCode also offers package managers. We lead with curl; from a checkout:

git clone https://github.com/agent-next/video-agent
cd video-agent
bash scripts/install.sh

Or download only: curl -fsSL https://open-video.ai/install.sh -o install.sh && bash install.sh --help

2. What the installer does

Detects OS, NVIDIA driver / CUDA, disk space, and tools.
Creates a Python venv and installs ComfyUI (the render engine).
Downloads MiniMax H3 INT8 ConvRot weights (~54 GB, resumable via aria2c).
Starts ComfyUI and optionally runs a short welcome generation.
Prints how to generate your next film from the CLI.

3. Generate a film

After install, from your open-video directory (the installer prints the path):

open-video "a neon koi swimming through rain, slow dolly"

Useful checks:

python cli/open_video.py list-models
python cli/open_video.py "anything" --dry-run

--dry-run validates the prompt and ComfyUI health without spending a full render.

4. VRAM cheat sheet (H3)

VRAM	Quant	Notes
~8 GB	NF4	Lowest entry; slower / softer quality
~10 GB	W4 ConvRot	Usable with care
~21 GB+	INT8 ConvRot (default)	Proven path (e.g. RTX 5090)
multi-GPU / huge	BF16	Full quality

Avoid NVFP4 on RTX 5090 (known ComfyUI issue). Prefer --lowvram when tight.

5. Try the director UI (no GPU)

The website Try page is a static mockup of the director loop (plan → generate → judge → refine). It plays the sample film so you can feel the flow. Real renders always run on your machine via the installer above.

6. Models & modes

See the model library for H3 (active) and upcoming backends (Wan, LTX). Modes: text-to-video, image-to-video / first-frame, first-last-frame.

Open Try UI Model library Home

In the open family

OpenVideo is designed next to sibling open products:

OpenCode — open coding agent (install UX + open-core playbook).
Open Design — open design workspace for local agents (desktop + skills + MCP).
OpenArt — creator-facing studio DNA (product, not node-graph).

Analogy: OpenCode → Cursor · Open Design → Claude Design · OpenVideo → Runway / Seedance. Related projects: repo docs/REFERENCES.md.

Troubleshooting

Install fails mid-download — re-run the same command; weights resume.
ComfyUI not reachable — confirm it is listening on 127.0.0.1:8188.
OOM / CUDA errors — use a lower quant, shorter duration, or --lowvram.
No NVIDIA GPU — you can browse the site and mock UI; generation backends are GPU-first today.

Questions? Source & issues on GitHub · site open-video.ai.