# LiveCanvas [简体中文](README.md) | **English** **Turn a sentence into a Live Photo with a designed cover and animation.** An agent skill for Codex, Claude Code, and other compatible assistants: **research → storyboard → Remotion animation → previews & Apple Live Photo resources**. Use images, emoji, scenes, timelines, and charts. ## Demos 15 animated GIF previews. Click any image to enlarge. [Downloads & notes](#downloads--notes) · [Install](#install) · [简体中文](README.md#作品演示)

Xiaomi market cap A20 Pro vs A19 Pro Introducing Gemini 4 Argon

1M output tokens On par with GPT-6.1 Sol Hallucination rate: 15%

#1 in Text Arena Terminal tasks: 57% Population

Coffee stores NEV adoption High-speed rail

Two-city temperatures Olympic gold Electricity mix

## Downloads & notes
Show descriptions, full-resolution MP4s and Live Photos ### Xiaomi market cap [Full-resolution MP4](docs/demos/en/xiaomi.mp4) ### A20 Pro vs A19 Pro [Full-resolution MP4](docs/demos/en/a20-pro.mp4) ### Introducing Gemini 4 Argon Engineering, knowledge work and security; phased access for trusted testers. [Full-resolution MP4](docs/demos/en/gemini-argon-intro.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-intro.pvt.zip) ### 1M output tokens The output limit rises from 64K to 1M tokens. [Full-resolution MP4](docs/demos/en/gemini-argon-output.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-output.pvt.zip) ### On par with GPT-6.1 Sol Artificial Analysis Intelligence Index: Argon 53, Astra 53, Sol 52. [Full-resolution MP4](docs/demos/en/gemini-argon-performance.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-performance.pvt.zip) ### Hallucination rate: 15% Artificial Analysis Omniscience: Argon 15%, Astra 51%, Sol 54%. [Full-resolution MP4](docs/demos/en/gemini-argon-hallucination.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-hallucination.pvt.zip) ### #1 in Text Arena 1525 points and 4,942 votes; ranked first, with preliminary results. [Full-resolution MP4](docs/demos/en/gemini-argon-arena.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-arena.pvt.zip) ### Terminal tasks: 57% Terminal-Bench 4.0: Argon 57%, versus 4% for Gemini 3.1 Pro Preview. [Full-resolution MP4](docs/demos/en/gemini-argon-terminal.mp4) · [Download Live Photo](docs/demos/en/gemini-argon-terminal.pvt.zip) ### Population [Full-resolution MP4](docs/demos/en/population.mp4) ### Coffee stores [Full-resolution MP4](docs/demos/en/coffee.mp4) ### NEV adoption [Full-resolution MP4](docs/demos/en/nev.mp4) ### High-speed rail [Full-resolution MP4](docs/demos/en/rail.mp4) ### Two-city temperatures [Full-resolution MP4](docs/demos/en/weather.mp4) ### Olympic gold [Full-resolution MP4](docs/demos/en/olympics.mp4) ### Electricity mix [Full-resolution MP4](docs/demos/en/energy.mp4) Data is a production-time snapshot, not a live feed. Gemini sources were checked on 2026-10-01; Text Arena is dated 2026-09-30. See [demo notes](docs/DEMOS.en.md) for sources and definitions.
## Install Requires Node.js / npm. Install with the [Skills CLI](https://github.com/vercel-labs/skills) and select your agent: ```bash npx skills add pengchujin/livecanvas --skill livecanvas -g ``` For Codex only: ```bash npx skills add pengchujin/livecanvas --skill livecanvas -g -a codex ``` Start a new chat after installation: ```text Use LiveCanvas to compare last year’s temperatures in Beijing and Shanghai. Create the content in English with city illustrations and seasonal emoji. Design a complete cover and deliver an animated preview. ```
Manual installation for Codex Clone or download this repository. Copy the entire `livecanvas/` folder to `~/.codex/skills/` (or `$CODEX_HOME/skills/` if configured), then start a new chat. Do not copy only SKILL.md. Copy the sibling `livechart/` folder too if you need the legacy `$livechart` alias.
## Requirements & output - **Animation:** An agent with web research and code execution; Node.js, Python 3, FFmpeg, and fonts for the content’s language. Install project dependencies with `npm ci`. - **Live Photo:** Pair Apple resources on Linux, Windows or macOS with Python 3.10+, Pillow and FFmpeg. macOS also offers a Swift backend and optional PhotoKit validation. Both accept silent H.264 video and JPEG images. See the [portable generation and iPhone import guide](livecanvas/references/portable-live-photo.md). - **Delivery:** Cover, MP4, sources, editable project, and a `.pvt` resource package where supported. No automatic Photos import. A `.pvt` bundles paired photo/video resources and metadata; iPhone Files may not preview it. README animations are GIFs, not native Live Photos. The template pins Remotion 4.0.530. See the [production guide (Chinese)](livecanvas/references/production.md) for manual rendering and packaging commands. ## License & credits Skill code and documentation are [MIT licensed](LICENSE). Demo media and third-party assets have [separate attribution and terms (Chinese)](THIRD_PARTY_NOTICES.md). Remotion retains [its own license](https://github.com/remotion-dev/remotion/blob/main/LICENSE.md). Motion design draws on [Emil Kowalski’s animate skill](https://github.com/emilkowalski/skills/blob/main/skills/animate/SKILL.md). Emoji graphics are from [Twemoji](https://github.com/twitter/twemoji).