AI-assisted rotoscoping in Google Colab. High-quality alpha mattes with no effort, saving you time.
All outputs are untouched — straight from the pipeline.
Difficulty: 🟦⬜⬜⬜⬜ Very Easy
Input Conditions: motion blur
demo_1.mp4
Difficulty: 🟩🟩⬜⬜⬜ Easy
Input Conditions: low contrast, color blending
demo_2.mp4
Difficulty: 🟨🟨🟨⬜⬜ Medium
Input Conditions: kinetic motion, frame-edge clipping, temporary object occlusion
demo_3.mp4
Difficulty: 🟧🟧🟧🟧⬜ Hard
Input Conditions: very low contrast, color blending, fine hair
demo_4.mp4
Difficulty: 🟥🟥🟥🟥🟥 Very Hard
Input Conditions: low contrast, multiple objects, objects overlapping
demo_5.mp4
- Click the Colab badge above to open the project
- Run both cells to install then launch
- Upload your footage and pick a tool
- Download your alpha sequence or composited result
That's it. No local setup. No terminal commands.
| Tool | Best For | How It Works |
|---|---|---|
| InSPyReNet | Salient subjects, good contrast | Fully automatic — no prompts needed |
| MatAnyone | More control, faster results | Single SAM prompt → auto-propagation |
| SAM + ViTMatte | Complex motion, low contrast | SAM tracking → edge refinement |
💡 Tip: Results vary by footage, A/B test to find the cleanest matte.
SAM + ViTMatte is typically the most consistent.
Full walkthrough from zero to compositing.
Click to watch on YouTube (3 min)
SAM 3 is included and fully functional, but the model weights require approval.
How to get access (5 min)
- Go to facebook/sam3 on Hugging Face
- Click "Agree and access repository" (requires HF account)
- Use AI researcher and a United States University for guaranteed form approval. (short wait)
- In Colab, paste your huggingface_token when prompted
MIT — use freely, attribution appreciated!
