RotoAI

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

(★ 59)

An open-source studio for prompt-driven video segmentation. Powered by SAM2 & Grounding DINO with a hybrid Cloud-Local architecture.

RotoAI Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 🟣 RotoAI - Intelligent Video Rotoscoping

<div align="center">
  
  ![RotoAI Banner](https://raw.githubusercontent.com/sPappalard/RotoAI/main/public/banner.png)
  
  [![License](https://img.shields.io/badge/License-AGPL_v3-blue.svg)](https://www.gnu.org/licenses/agpl-3.0)
  [![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
  [![React 18](https://img.shields.io/badge/react-18.0+-61dafb.svg)](https://reactjs.org/)
  [![FastAPI](https://img.shields.io/badge/FastAPI-0.104+-009688.svg)](https://fastapi.tiangolo.com/)
  [![Powered By SAM2](https://img.shields.io/badge/AI-SAM2%20%2B%20GroundingDINO-purple)](https://github.com/facebookresearch/segment-anything-2)
  
  **Automated zero-shot video segmentation powered by SAM2 & Grounding DINO**
  
  [🌟 Overview](#-overview) • 
  [🎨 Visual Effects Showcase](#-visual-effects-showcase) •
  [✨ Key Features](#-key-features) •
  [📸 UI Screenshots](#-ui-screenshots) •
  [🔄 Pipeline & Architecture](#-how-it-works) •  
  [🛠 Tech Stack & Repository Structure](#-tech-stack) •
  [💾 Memory Management](#-Memory-Management) •
  [🚀 Getting Started](#-getting-started) • 
  [🐛 Troubleshooting](#-troubleshooting) •  
  [📜 Credits](#-Credits--Acknowledgments) •
  [👨‍💻 Author](#author)
  <br><br>

**If you find RotoAI useful, please consider supporting the development!** 
  
  <a href="https://www.buymeacoffee.com/sPappalard">
    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" width="180" />
  </a>
  
</div>

---

## 🌟 Overview

**RotoAI** is an advanced open-source studio for prompt-driven video segmentation. 
It leverages a **Hybrid Cloud-Local Architecture**: a responsive React frontend runs locally, while the heavy inference is offloaded to **Google Colab GPUs** (T4) via a secure Ngrok tunnel.
Powered by state-of-the-art foundation models (**SAM2** & **Grounding DINO**), RotoAI introduces intelligent VRAM management and chunked processing, enabling high-resolution rotoscoping on free cloud tier hardware without memory bottlenecks.

### What Makes RotoAI Special?

- **Semantic Understanding**: Select objects using natural language prompts (e.g., *"person in red shirt"*) via Grounding DINO.
- **Hybrid Architecture**: Combines the responsiveness of a local UI with the raw power of Cloud GPUs (Google Colab).
- **Production Resilience**: Handles long videos via **Smart Chunking** (5s segments) and **Auto-Resolution Scaling** to prevent OOM errors.
- **Dual Detection Modes**: Supports both generic Zero-Shot detection and **Custom YOLO Models** for specialized tasks.
- **6 Professional Effects**: From cinematic B&W pop to neon glow overlays

---

## 🎬 Demo

<div align="center">
  
  **Watch RotoAI in Action**
  
  [![RotoAI Demo](https://img.youtube.com/vi/bGGD_3j_rJY/maxresdefault.jpg)](https://www.youtube.com/watch?v=bGGD_3j_rJY)
  
  *Click to watch the full demonstration on YouTube*
  
</div>

---

## 🎨 Visual Effects Showcase

Discover the cinematic effects you can create in seconds.

### Bokeh Blur
*Simulates a high-end camera lens by applying a realistic Gaussian blur to the background, creating a shallow depth-of-field effect.*
<div align="center">
  <img src="public/boke.gif" alt="Bokeh Blur Effect" width="800">
  <p>Prompt used: <em>"Running man"</em></p>
</div>

---

### Chroma Key (Green Screen)
*Replaces the background with a solid green (or custom hex) color, perfect for compositing in post-production tools like After Effects or Premiere.*
<div align="center">
  <img src="public/greenk.gif" alt="Chroma Key Effect" width="800">
  <p>Prompt used: <em>"Boys dressed in red"</em></p>
</div>

---

### B&W Color Pop
*Isolates the subject by keeping them in full color while instantly desaturating the background to grayscale.*
<div align="center">
  <img src="public/bew.gif" alt="B&W Effect" width="800">
  <p>Prompt used: <em>"Boy with orange backpack"</em></p>
</div>

---

### Neon Glow
*Adds a futuristic glowing outline around the detected subject. You can choose between a sharp border or a diffuse glow.*
<div align="center">
  <img src="public/neonEffect.gif" alt="Neon Effect" width="800">
  <p>Prompt used: <em>"Dancing man"</em></p>
</div>

**Configuration Options**:
- ✅ With Border: Colored neon outline with edge detection
- ❌ No Border: Soft glow with adjustable blur radius (1-15)

---

### Color Pop
*Applies a cinematic desaturation filter to the background, creating a moody, vintage aesthetic while keeping the subject vivid.*
<div align="center">
  <img src="public/colorPop.gif" alt="Color Pop Effect" width="800">
  <p>Prompt used: <em>"Man with glasses"</em></p>
</div>

---

### Luminous Edge
*Highlights the contours of the subject with a radiant light effect, creating a sketched light-painting look.*
<div align="center">
  <img src="public/doctors.gif" alt="Edge Light Effect" width="800">
  <p>Prompt used: <em>"Doctors"</em></p>
</div>

---

## ✨ Key Features

### 🤖 AI-Powered Detection
- **Open-Vocabulary Detection**: Zero-shot capabilities via **Grounding DINO** allows finding any object using natural language prompts.
- **BYO Model (Bring Your Own)**: Support for custom trained **YOLO (.pt)** weights for specialized industrial or specific object detection.
- **Interactive Calibration**: Built-in **Test Mode** to validate detection accuracy on individual frames before committing to full GPU rendering.

### 🎨 Professional Visual Effects

**Cinematic Effects**
- 🔳 **B&W Color Pop**: Isolate subjects in vibrant color against grayscale backgrounds
- 🟩 **Chroma Key**: Green screen-style background replacement with custom colors
- 🧪 **Neon Glow**: Cyberpunk-inspired luminous edge effects with configurable colors

</td>
<td width="50%">

**Advanced Filters**
- 💧 **Bokeh Blur**: Professional depth-of-field simulation
- 🎞️ **Color Pop**: Cinematic desaturation for mood creation
- 💡 **Luminous Edge**: Highlight subject contours with glowing borders

</td>
</tr>
</table>

### ✨Effect Previews

<p><i>▶️ Click play to watch the Effect Previews</i></p>
</div>

---

### ⚙️ Advanced Configuration
- **Dual Output Modes**: Side-by-side comparison or processed-only
- **Smart Scanning**: Configurable detection window (1-10 seconds)
- **Precision Tuning**: Adjustable confidence thresholds (0.01-0.80)
- **Memory Management**: Automatic resolution scaling for optimal VRAM usage

### 🚀 Performance & Optimization
- **Chunked Processing**: Handles videos of any length without OOM errors
- **GPU Acceleration**: FP16 precision for 2x speed on modern GPUs
- **Real-Time Progress**: Frame-by-frame statistics with ETA
- **Smart Caching**: Efficient frame storage and cleanup

---

## 📸 UI Screenshots

A glance at the RotoAI interface and its capabilities.

### Main interface

### Hybrid Cloud Connection

The entry point connecting the local React UI with the Colab GPU backend via Ngrok.

### Interactive test detection

**Test Detection Module:** Validate prompts and confidence thresholds frame-by-frame before processing.

### Visual Effects Engine

Select from cinematic effects like **Neon Glow**, **Chroma Key**, **Bokeh Blur**, and **B&W Pop**.

### Advanced Configuration

Fine-tune scan duration, confidence thresholds, and output formats (Comparison vs. Processed Only).

### Results & Player

Built-in video player with loop functionality and instant download for the final rendered MP4.

---

## 🔄 How It Works

### AI Processing Pipeline

<div align="center">
  
  ![RotoAI Pipeline](https://raw.githubusercontent.com/sPappalard/RotoAI/main/public/pipeline.png)
  
  *End-to-End Flow: Visual step-by-step of the rendering process*
  
</div>

### Hybrid Architecture

<div align="center">
  
  ![RotoAI Architecture](https://raw.githubusercontent.com/sPappalard/RotoAI/main/public/architecture.png)
  
  *Hybrid Infrastructure: Local Frontend connected to Remote Backend via Ngrok*
  
</div>

---

## 🛠 Tech Stack

### Core AI Models

### 🔍 Detection (The "Eyes")
> **Grounding DINO (SwinB)**
> *The prompt-master.* Allows you to select objects using natural language queries (e.g., "black cat").
> * **Size:** ~600MB
> * **Type:** Zero-shot Object Detection

> **YOLO v8/v11**
> *The specialist.* Supports user-uploaded `.pt` weights for fine-tuned tasks.
> * **Size:** <100MB (Typical)
> * **Type:** Custom Object Detection

### ✂️ Segmentation (The "Hands")
> **SAM 2 (Segment Anything 2)**
> *The tracker.* Handles the heavy lifting of propagating masks across video frames.
> * **Architecture:** Hiera Small
> * **Size:** ~180MB
> * **Performance:** Real-time propagation

### Backend Stack

```python
# Core Dependencies
PyTorch 2.0+          # Deep learning framework
FastAPI 0.104+        # Async web framework
Uvicorn               # ASGI server
OpenCV (cv2)          # Video processing
NumPy                 # Matrix operations
Pillow (PIL)          # Image manipulation
FFmpeg                # Video encoding
```

**Infrastructure**:
- 🌐 **Google Colab**: Cloud GPU environment (T4/P100/V100)
- 🔗 **Ngrok**: Secure tunnel for public API access
- 🔧 **Nest AsyncIO**: Event loop management for Colab

### Frontend Stack

```javascript
// Core Technologies
React 18              // UI library
Next.js 14            // React framework
Vite 5                // Build tool
Tailwind CSS 3        // Utility-first styling
Lucide React          // Icon set
```

## 📂 Repository Structure (Codebase)

This directory contains the source code for the UI and the Backend logic.

> **Note:** Large model weights are excluded from the repository to keep it lightweight. They are handled separately via Google Drive.

```text
RotoAI/
│
├── be/                                # Backend (Python & Colab)
│   ├── app.py                         # FastAPI Entry Point (Inference Server)
│   └── setup_colab.ipynb              # Setup Script for Google Colab
│
├── fe/                                # Frontend (Next.js + TypeScript)
│   ├── app/                           # Next.js App Router
│   │   ├── globals.css                # Global Styles
│   │   ├── layout.tsx                 # Root Layout
│   │   └── page.tsx                   # Main Application Page
│   │
│   ├── public/                        # Static Assets (Logos, Icons)
│   ├── next.config.ts                 # Next.js Configuration
│   ├── postcss.config.mjs             # PostCSS Config
│   ├── package.json                   # Frontend Dependencies
│   └── tsconfig.json                  # TypeScript Configuration
│
├── public/                            # Demo media and visual documentation
├── .gitignore                         # Git Ignore Rules
└── README.md                          # Project Documentation
```

## ☁️ Google Drive Integration (Persistent Model Cache)

Since Google Colab environments are ephemeral (data is lost when you disconnect), RotoAI automatically creates a dedicated folder in your Google Drive during the first setup.

```text
MyDrive/
└── RotoAI_Models/                       # Created automatically by setup_colab.ipynb
    │
    ├── sam2_hiera_small.pt              # SAM 2 Weights (~180MB)
    │                                    # Downloaded once, reused forever.
    │
    ├── groundingdino_swinb_cogcoor.pth  # Grounding DINO Weights (~600MB)
    │                                    # Prevents large downloads on every boot.
    │
    └── GroundingDINO_SwinB_cfg.py       # Model Configuration
                                         # Auto-copied from the repo for compatibility.
```

---

### 💾 Memory Management

#### Automatic Resolution Scaling

RotoAI intelligently manages VRAM to prevent crashes:

```python
# Backend logic (simplified)
def calculate_scale_factor(video_width, video_height, total_frames):
    """
    Estimates VRAM needed and scales resolution if necessary.
    """
    # Memory needed (bytes)
    needed = width * height * 3 * total_frames * 4  # float32
    needed_gb = needed / (1024**3) + 3.0  # +3GB for models
    
    # Memory available
    total_vram = torch.cuda.get_device_properties(0).total_memory / (1024**3)
    used_vram = torch.cuda.memory_allocated() / (1024**3)
    available = total_vram - used_vram
    
    # Scale factor
    if needed_gb > available:
        scale = max(0.5, min(0.9, available / needed_gb * 0.85))
    else:
        scale = 1.0  # No scaling needed
    
    return scale
```

**Example Scenarios**:

| Video Resolution | Total Frames | VRAM | Scale Factor | Processed Resolution |
|------------------|--------------|------|--------------|---------------------|
| 1920x1080 (FHD) | 900 (30s) | 16GB | 1.0 | 1920x1080 |
| 3840x2160 (4K) | 1800 (60s) | 16GB | 0.5 | 1920x1080 |
| 1920x1080 (FHD) | 5400 (3min) | 8GB | 0.7 | 1344x756 |

**Key Insights**:
- ⚠️ **Never scales below 50%** (maintains quality)
- ⚠️ **Never scales above 90%** (safety margin)
- ✅ **Dimensions aligned to 16** (codec requirement)

#### Chunked Processing

Videos are split into 5-second chunks:

```python
chunk_duration = 5  # seconds
chunk_frames = int(fps * chunk_duration)  # e.g., 30fps * 5s = 150 frames

for chunk_idx in range(num_chunks):
    # 1. Extract frames for this chunk
    frames = extract_chunk(video, start_frame, end_frame)
    
    # 2. Initialize SAM2 for chunk
    sam2.init_state(frames)
    
    # 3. Propagate masks frame-by-frame
    for frame_idx, mask in sam2.propagate():
        rendered_frame = apply_effect(frame, mask, effect_type)
        save(rendered_frame)
    
    # 4. Cleanup (free VRAM)
    sam2.reset_state()
    del frames
```

**Why 5 Seconds?**:
- ✅ Balances memory usage vs. tracking continuity
- ✅ ~150 frames at 30fps (manageable for SAM2)
- ✅ Smooth transitions between chunks

---

## 🚀 Getting Started

Follow these steps to set up the **Hybrid Architecture**:

1. **☁️ Cloud:** Start the heavy AI backend on Google Colab.
2. **💻 Local:** Start the user interface on your machine.
3. **🔗 Connect:** Link them together via Ngrok.

### Prerequisites

* **Google Account:** To access [Google Colab](https://colab.research.google.com).
* **Ngrok Account:** [Sign up for free](https://ngrok.com) and get your **Auth Token**.
* **Local Machine:** Node.js (v18+) and Git installed.

---

### Phase 1: Backend Setup (Google Colab)

The backend handles the heavy AI processing. We run this on Google's servers to save your local resources.

#### 1. Prepare the Notebook

1. Go to [Google Colab](https://colab.research.google.com).
2. Upload the file `be/setup_colab.ipynb` from this repository (or create a new notebook).
3. **Crucial:** Go to menu **Runtime** → **Change runtime type** → select **T4 GPU**.

#### 2. Set API Keys

1. Click the **Key Icon (Secrets)** 🔑 in the left sidebar of Colab.
2. Add a new secret:
* **Name:** `NGROK_TOKEN`
* **Value:** `Your_Ngrok_Auth_Token` (from your dashboard).

3. Toggle the "Notebook access" switch to **ON**.

#### 3. Install & Run

Run the cells in the notebook in order. They will:

1. **Verify GPU** (Must be T4).
2. **Mount Google Drive** (To create the `RotoAI_Models` cache folder).
3. **Install Dependencies** (SAM2, GroundingDINO, etc.).
4. **Start the Server**.

At the end, you will see a log like this:

```text
🚀 SERVER v7.2 ONLINE!
🔗 https://a1b2-34-56-78-90.ngrok-free.app

```

> 📋 **Copy this URL.** You will need it for the frontend.

---

### Phase 2: Frontend Setup (Local Machine)

The frontend is the visual interface running on your computer.

#### 1. Clone & Install

Open your terminal and run:

```bash
# 1. Clone the repository
git clone https://github.com/sPappalard/RotoAI.git

# 2. Navigate to the frontend folder
cd RotoAI/fe

# 3. Install dependencies
npm install

```

#### 2. Launch the App

Start the development server:

```bash
npm run dev

```

Your terminal will show:

```text
  ▲ Next.js 14.2.3
  - Local: http://localhost:3000

```

Open **http://localhost:3000** in your browser.

---

### Phase 3: Connect & Create

1. On the web page (`localhost:3000`), you will see a **"Backend Connection"** input field.
2. Paste the **Ngrok URL** you copied from Colab (Step 1).
3. Click **Connect**.

✅ **You are ready!** The status indicator should turn **Green**.

#### ⚡ Quick Usage Guide

1. **Upload:** Drag & drop a short video (recommended: 10-15s for testing).
2. **Prompt:** Type what you want to rotoscope (e.g., *"black dog"*, *"person dancing"*).
3. **Effect:** Choose a visual effect (e.g., *Neon, Green Screen*).
4. **Generate:** Click "Generate Roto Effect" and watch the progress bar!

---

## 🐛 Troubleshooting

### Common Issues & Solutions

#### Backend Issues

<details>
<summary><b>🔴 "NGROK_TOKEN not found" error</b></summary>

**Problem**: Ngrok authentication token not configured in Colab Secrets.

**Solution**:
1. Go to https://dashboard.ngrok.com/get-started/your-authtoken
2. Copy your token
3. In Colab: Click 🔑 icon (left sidebar) → "Add a new secret"
4. Name: `NGROK_TOKEN`
5. Value: Paste your token
6. Toggle switch to enable
7. Restart Colab runtime

</details>

<details>
<summary><b>🔴 "CUDA out of memory" error</b></summary>

**Problem**: Video too large for available VRAM.

**Solutions**:
1. **Use shorter test clip** (10-15 seconds)
2. **Restart Colab runtime** to free memory
3. **Try lower resolution source** (720p instead of 1080p)
4. The app *should* auto-scale, but manual scaling may help:
   ```python
   # In app.py, force scaling
   scale = 0.7  # 70% resolution
   ```

**Note**: T4 GPU (Colab free tier) has 16GB VRAM. For 4K videos >1 minute, consider Colab Pro (V100, 40GB).

</details>

<details>
<summary><b>⚠️ "No object found in scan period"</b></summary>

**Problem**: Detection failed to find objects in initial frames.

**Solutions**:
1. **Increase scan duration**: Advanced Settings → Scan Duration → 8-10s
2. **Lower confidence threshold**: 0.35 → 0.25
3. **Try different prompt**: 
   - Bad: `"person walking"` (too specific)
   - Good: `"person"` (simple, general)
4. **Test Detection first**: Verify object is visible in scanned frames

</details>

<details>
<summary><b>🔴 Server disconnected during processing</b></summary>

**Problem**: Colab runtime timed out or Ngrok tunnel closed.

**Causes**:
- Colab inactivity (90 minutes free tier)
- Ngrok free tier session limit (2 hours)
- Browser tab closed

**Solutions**:
1. **Keep Colab tab active** (interact every 60 minutes)
2. **Use Colab Pro** for longer sessions
3. **Upgrade Ngrok** for permanent tunnels
4. **Split long videos** into shorter segments

</details>

#### Frontend Issues

<details>
<summary><b>🔴 "Failed to upload video. Please check URL."</b></summary>

**Problem**: Frontend can't reach backend API.

**Checklist**:
1. ✅ Is backend cell running in Colab?
2. ✅ Is Ngrok URL copied correctly? (no trailing slash)
3. ✅ Does URL start with `https://`?
4. ✅ Try visiting URL in new tab—should show FastAPI docs
5. ✅ Check browser console for CORS errors

**Test Connectivity**:
```bash
# In terminal
curl https://your-ngrok-url.ngrok-free.app/status
# Should return: {"status":"idle","progress":0,...}
```

</details>

<details>
<summary><b>⚠️ Video player shows black screen</b></summary>

**Problem**: Processed video failed to load or is corrupted.

**Solutions**:
1. **Check format**: Only MP4/H.264 supported in browser
2. **Re-download**: Right-click video → "Save video as..."
3. **Try VLC**: Open downloaded file in VLC Player to verify
4. **Re-process with different settings**

</details>

<details>
<summary><b>🔴 "Test Detection" stuck loading frame</b></summary>

**Problem**: Backend not responding to `/extract-frame` endpoint.

**Solutions**:
1. **Check Colab**: Is runtime still active?
2. **Refresh page**: Sometimes state gets stale
3. **Re-upload video**: Upload process may have failed silently
4. **Check total_frames**: Should be >0 after upload

**Debug**:
```javascript
// In browser console
console.log(totalFrames);  // Should be >0
```

</details>

---

## 📜 Credits & Acknowledgments

### Core Technologies

- **[SAM2](https://github.com/facebookresearch/segment-anything-2)** by Meta AI Research
  - *State-of-the-art video segmentation model*
  - Paper: "Segment Anything in Images and Videos" (Kirillov et al., 2024)

- **[Grounding DINO](https://github.com/IDEA-Research/GroundingDINO)** by IDEA-Research
  - *Zero-shot object detection with text prompts*
  - Paper: "Grounding DINO: Marrying DINO with Grounded Pre-Training" (Liu et al., 2023)

- **[YOLO](https://github.com/ultralytics/ultralytics)** by Ultralytics
  - *Fast and accurate object detection framework*
  - Used for custom model support

### Frameworks & Libraries

- **[FastAPI](https://fastapi.tiangolo.com/)** - Modern Python web framework
- **[React](https://reactjs.org/)** - JavaScript UI library
- **[Tailwind CSS](https://tailwindcss.com/)** - Utility-first CSS framework
- **[PyTorch](https://pytorch.org/)** - Deep learning framework
- **[OpenCV](https://opencv.org/)** - Computer vision library
- **[FFmpeg](https://ffmpeg.org/)** - Multimedia processing

### Special Thanks

- **Google Colab** - Free GPU compute for AI researchers
- **Ngrok** - Secure tunneling for local development
- **Roboflow** - Computer vision development platform

---

<a name="author"></a>
## 👨‍💻 Author

**RotoAI** is created and maintained by **sPappalard**.

If you find this project useful, please give it a ⭐ star or support the development!

[![GitHub](https://img.shields.io/badge/GitHub-sPappalard-181717?style=for-the-badge&logo=github)](https://github.com/sPappalard)
[![LinkedIn](https://img.shields.io/badge/LinkedIn-Connect-0A66C2?style=for-the-badge&logo=linkedin)](https://www.linkedin.com/in/salvatore-pappalardo98/)
[![Email](https://img.shields.io/badge/Email-Contact_Me-D14836?style=for-the-badge&logo=gmail&logoColor=white)](mailto:salvatore.pappalardo.job@gmail.com)  
[![Buy Me A Coffee](https://img.shields.io/badge/Buy_Me_A_Coffee-FFDD00?style=for-the-badge&logo=buy-me-a-coffee&logoColor=black)](https://www.buymeacoffee.com/spappalard)

</div>

---

## 📄 License

This project is released under the **GNU Affero General Public License v3.0 (AGPL-3.0)**.

**Why AGPL-3.0?**
This project integrates **Ultralytics YOLO**, which is licensed under AGPL-3.0. As a derivative work, RotoAI inherits this license to ensure full compliance with the open-source terms of its dependencies.

**What this means for you:**
- ✅ **Use:** You can use this software for personal, research, or commercial purposes.
- ✅ **Modify:** You can modify the source code.
- 🔄 **Share:** If you distribute this software or host it as a network service (SaaS), you **must** disclose the source code of your modified version under the same AGPL-3.0 license.

---
*Note: This project also utilizes SAM 2 (Apache 2.0) and Grounding DINO (Apache 2.0).*

<div align="center">
  
  **Built with ❤️ by [@sPappalard](https://github.com/sPappalard)**
  
  <br>
  
  [⬆ Back to Top](#readme-top)
  
</div>

RotoAI

RotoAI Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

RotoAI

RotoAI Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README