📺 Watch the Vox AI Motion Graphics Generator in action →
🚨 The problem with making explainer videos today:
- ❌ Fragmented workflow — separate tools for scripting, image gen, animation, voice-over, music, and captions
- ❌ No narrative structure — raw text-to-video models don’t understand story arcs, hooks, or pacing
- ❌ Inconsistent look — the collage aesthetic drifts from shot to shot
- ❌ Hours of manual editing — stitching, ducking music, and burning captions by hand
💡 The solution:
🎬 Screenwriter, Collage Artist, Animator, and Editor — all in one automated pipeline.
Type one topic. The agent writes the story beats, renders each beat as a torn-paper collage poster, animates it, adds a narrator voice-over, music, and burned-in captions — then stitches everything into a finished final.mp4. You stay in control with just two approval gates: the story beat map and the visual theme.
🎨 Explore 50+ more open-source AI apps →
🎥 Demo
https://github.com/user-attachments/assets/3db826d1-3271-4e8a-b4d0-a843013e67c7
One topic in, a fully narrated & captioned Vox-style collage video out. (download the demo)
“Make me a 15-second Vox-style collage video on the history of coffee.” → a styled final.mp4, fully narrated and captioned.
📽️ Generated Showcase Videos (Playable Demo Videos)
Below are end-to-end videos generated using the pipeline:
| Mode |
Film Title |
Topic / Subject |
Playable Video Link |
Keyframe Poster |
Motion Model |
Narration Voice |
| B-roll |
☕ History of Coffee |
Ethiopia 850 AD |
Play Video (out/demo/final.mp4) |
out/demo/keyframes/kf_1a.jpg |
veo3.1-image-to-video |
MiniMax TTS (Q19bea09caa6IRAeW7) |
| C-roll |
🍾 Cold Brew Product Ad |
Artisanal Cold Brew Bottle |
Play Video (out/demo-croll/final.mp4) |
out/demo-croll/keyframes/kf_1.jpg |
veo3.1-image-to-video |
MiniMax TTS (Q19bea09caa6IRAeW7) |
🎬 Generated Assets Summary:
- ☕ B-roll (History of Coffee):
- 📹 Video: Play Video (
out/demo/final.mp4) | 🖼️ Poster: out/demo/keyframes/kf_1a.jpg | 📹 Clip: out/demo/clips/clip_1a.mp4 | 🎙️ Voice: out/demo/audio/narr_1.mp3 | 🎵 BGM: out/demo/audio/bgm.mp3
- 🍾 C-roll (Cold Brew Product Anchor):
- 📹 Video: Play Video (
out/demo-croll/final.mp4) | 📸 Anchor Photo: out/demo-croll/anchor_photo.jpg | 🖼️ Poster: out/demo-croll/keyframes/kf_1.jpg | 📹 Clip: out/demo-croll/clips/clip_1.mp4 | 🎙️ Voice: out/demo-croll/audio/narr_1.mp3 | 🎵 BGM: out/demo-croll/audio/bgm.mp3
📑 Table of Contents
✨ Key Features
- 3 Input Modalities (A-roll, B-roll, C-roll):
- B-roll (Topic → Film) — a single one-line prompt produces a complete, captioned, narrated video
- A-roll (Talking-Head Video → Collage) — segment a presenter video with ASR (
openai-whisper) and re-style into paper-collage while keeping real face, gestures, and lip-sync
- C-roll (Single Photo / Product → Anchored Collage) — cut out a selfie or product shot as a photographic sticker and build animated collage posters around it
- Authentic Vox collage aesthetic — torn paper, cutouts, tape, halftone dots, newspaper clippings, bold flat color, big headlines
- Story-first — picks a narrative arc (timeline, problem-agitate-solution, how-it-works…) and writes a hook-led beat map before generating anything
- Style bake-off — renders the same beat in 3–4 themes so you pick the look before committing
- Living-poster motion — animates each collage keyframe into dynamic motion with Google Veo 3.1
- Voice + music + captions — narration (TTS or voice cloning), background music ducked under the voice, and burned-in captions, all automated
- Two human gates, everything else automated — approve the beat map, pick the theme; the pipeline handles the rest
- Agent-native — a self-contained skill any coding agent (Claude Code, Codex, Antigravity, …) can read and run
- One API key + ffmpeg — no cluster of accounts to wire up
🎨 The Look
The aesthetic is the modern editorial paper-collage popularized by Vox explainers and creators like Stav Zilber and rom1trs: hand-cut paper cutouts, torn edges, tape, halftone dots, newspaper clippings, bold flat colors per beat, and big cutout headlines — brought to life with dynamic motion, a narrator voice-over, music, and burned-in captions.
The collage look is born in the image step (each beat is a finished collage poster), and the motion is added after — so the DNA of the style is locked in before anything moves.
🔄 How It Works
A single topic, video, or photo flows through the pipeline driven by beats.json under out/<project>/:
topic / video / photo
│
├─ 1. Beat Map Pick a narrative arc → write beats.json ◀── GATE 1: approve the beat map
├─ 2. Style Bake-Off Render the same beat in 3–4 themes ◀── GATE 2: pick the look
├─ 3. Keyframes One collage poster per beat (nano-banana-2 / flux-dev)
├─ 4. Motion Animate each poster into a clip (veo3.1-image-to-video)
├─ 5. Voice & Music Narration (minimax-speech-2.6) + BGM (suno-create-music)
├─ 6. Assemble Stitch clips, duck music, burn captions (ffmpeg)
└─ final.mp4
3 Modalities:
- B-roll (Topic → Film): Topic in → script → keyframe posters → Veo 3.1 motion → TTS narration + Suno music →
final.mp4.
- A-roll (Talking-Head → Collage): Presenter video in → Whisper ASR beat segmentation → video-edit restyling → audio remuxing →
final.mp4.
- C-roll (Photo / Product → Collage): Product photo in → photographic sticker cutout → anchored collage posters → Veo 3.1 motion → cloned voice →
final.mp4.
🧩 Models
| Pipeline Job |
Model Endpoint |
| Keyframes / Collage Posters |
nano-banana-2 / flux-dev |
| Motion / Animation |
veo3.1-image-to-video / veo3.1-fast-image-to-video / runway-image-to-video |
| Re-style Talking-Head (A-roll) |
gemini-omni-video-edit / veo3.1-image-to-video |
| Anchor Photo in Collage (C-roll) |
nano-banana-2 / flux-dev |
| Narration (TTS) |
minimax-speech-2.6-turbo |
| Voice Cloning (C-roll) |
minimax-voice-clone |
| Background Music |
suno-create-music |
| Audio Transcription (A-roll ASR) |
openai-whisper |
| Background Removal |
remove-background |
🚀 Quick Start
1. Install local dependencies
- ffmpeg + ffprobe (
brew install ffmpeg on macOS / choco install ffmpeg on Windows)
- Python 3 with Pillow (
pip install pillow)
2. Configure environment keys
export MUAPI_API_KEY="your-api-key" # image / video / voice / music models — key from muapi.ai
export OPENAI_API_KEY="your-openai-key" # story planning & ASR
3. Run Pipeline via Agent or CLI:
Mode 1: B-roll (Topic → Film)
Ask your coding agent:
“Make me a 15-second Vox-style collage video introducing the history of coffee.”
Or run CLI scripts manually:
python scripts/style_bakeoff.py out/my-topic american-retro,swiss-modern,punk-zine
python scripts/keyframes.py out/my-topic
python scripts/clips.py out/my-topic
python scripts/audio.py out/my-topic
python scripts/assemble.py out/my-topic
Mode 2: A-roll (Talking-Head Video → Collage)
python scripts/asr_beats.py out/my-aroll source_presentation.mp4
python scripts/aroll_clips.py out/my-aroll
python scripts/aroll_assemble.py out/my-aroll
Mode 3: C-roll (Single Photo / Product → Anchored Collage)
python scripts/croll_keyframes.py out/my-croll
python scripts/clips.py out/my-croll
python scripts/audio.py out/my-croll
python scripts/assemble.py out/my-croll
🛠️ Project Structure
.
├── SKILL.md # full agent workflow + approval gates
├── AGENTS.md # entry point for coding agents (Claude Code, Codex, Antigravity)
├── README.zh.md # Chinese documentation
├── scripts/ # one script per pipeline stage
│ ├── muapi_client.py # MuAPI client wrapper
│ ├── provider.py # task queue & status polling
│ ├── style_bakeoff.py # multi-theme style candidates
│ ├── keyframes.py # B-roll collage poster generator
│ ├── clips.py # B-roll video animator (Veo 3.1)
│ ├── audio.py # TTS narration & Suno music generator
│ ├── text_overlay.py # subtitle captions & watermark generator
│ ├── assemble.py # ffmpeg video assembly, ducking & composition
│ ├── asr_beats.py # A-roll Whisper ASR beat cutter
│ ├── aroll_clips.py # A-roll talking-head video restyler
│ ├── aroll_assemble.py # A-roll clip & audio remuxer
│ └── croll_keyframes.py # C-roll photo sticker anchor generator
├── references/ # prompt guide, beat/story library, voices
└── examples/ # sample beats.json
This is an agent skill — Claude Code auto-loads it from SKILL.md; Codex and other agents follow SKILL.md via AGENTS.md. Just ask for a “vox video” or a “collage video.”
评论