VOID
#21 v AI úprava videanetflix · od 2. April 2026 (arXiv-Paper); Code/Gewichte auf GitHub und Hugging Face ab ca. 2.–4. April 2026 öffentlich verfügbar · 21× · naposledy 30. 6. 2026
VOID (Video Object and Interaction Deletion) is an open-source AI model developed by Netflix researchers in collaboration with INSAIT Sofia University for removing objects from video. Unlike conventional video inpainting tools, VOID not only erases the target object and surface-level effects like shadows or reflections, but also detects and corrects the physical downstream consequences of the removal, such as objects falling once a supporting person is deleted. The model is built on Alibaba's CogVideoX-Fun diffusion model, uses a quadmask conditioning scheme (four mask values instead of a binary mask), and was trained on synthetic, physically simulated datasets (Kubric, HUMOTO). Code and weights were released under the Apache 2.0 license on GitHub and Hugging Face in April 2026, including
Vlastnosti
| Output Formats | Video-to-video (MP4); resolution 384x672, up to approx. 197 frames |
| Base Model | CogVideoX-Fun-V1.5-5b-InP (Alibaba PAI, 5B parameter video diffusion model), fine-tuned with interaction-aware quadmask conditioning |
| Integrations | Uses SAM2 (Meta) for segmentation and Gemini 3 Pro (Google) for VLM scene analysis; base CogVideoX-Fun-V1.5-5b-InP (Alibaba PAI) |
| Collaboration | Community adoption via GitHub/Hugging Face; public Gradio demo (Space: sam-motamed/VOID) for testing without own hardware |
| License | Apache License 2.0 (code and model weights, commercial use permitted) |
| Platform | GitHub (netflix/void-model), Hugging Face Model Hub + Gradio demo space, Colab notebook; local via Python/CLI |
| Price | Free (open source, no fees) |
| Release Date | arXiv paper April 2, 2026; code/model weights released approx. April 2–4, 2026 |