Raw unedited outputs.
demo_1.mp4
🟦⬜⬜⬜⬜ Very Easy: motion blur
demo_2.mp4
🟩🟩⬜⬜⬜ Easy: low contrast, outfit blending
demo_3.mp4
🟨🟨🟨⬜⬜ Medium: fast motion, subject and objects leaving frame
demo_4.mp4
🟧🟧🟧🟧⬜ Hard: very low contrast, outfit blending, fine hair
demo_5.mp4
🟥🟥🟥🟥🟥 Very Hard: low contrast, overlapping subjects
- Click Open In Colab above
- Run Cell 1 to install, then Cell 2 to launch
- Upload your video and choose a tool
- Download your exported alpha sequence
That's it. No local setup or terminal commands.
| Tool | Best For | Workflow |
|---|---|---|
| InSPyReNet | Hands-off cuts for clear subjects | Upload video and click Run |
| MatAnyone | Quick results with manual control | Click 1 frame, then click Run |
| SAM + ViTMatte | Complex motion, hair, soft edges | Click subject, Track, run ViTMatte |
💡 Results vary by footage. Try multiple tools to find the cleanest matte. SAM + ViTMatte is usually the most consistent.
Click to watch on YouTube (3 min)
SAM 3 is supported, but Meta requires approval before downloading the model weights.
How to get access
- Go to facebook/sam3 on Hugging Face
- Click Agree and access repository (free account required)
- Select AI researcher and a United States University for fast approval
- In Colab, provide your Hugging Face token when prompted
MIT. Use freely, attribution appreciated!
