UniVideo is an open source unified AI model that handles video understanding, generation (from text or images), and editing all within one framework using natural language instructions.

Is UniVideo free and open source?

Yes, it's completely free and open source. Model weights and inference code are available on Hugging Face (KlingTeam/UniVideo) and GitHub (KwaiVGI/UniVideo).

When was UniVideo released?

The research paper was published on October 9, 2025 (arXiv:2510.08377), with model and code released shortly after for community use.

What makes UniVideo different from other video AI tools?

Its dual stream design (MLLM + MMDiT) allows true unification of tasks, strong zero shot editing generalization (e.g., green screening without training), task composition, and visual-prompt support, outperforming many task specific models in benchmarks.

UniVideo

One Model for Video Understanding, Generation, and Editing

From KwaiVGI (Kuaishou)

Founder: Cong Wei, Quande Liu, Zixuan Ye, Qiulin Wang, Xintao Wang, Pengfei Wan, Kun Gai, Wenhu Chen

Video & Animation

Oct 2025

10K+

Free

Pricing Model

Free

Starting Price

$0/Month

About This AI

UniVideo is an advanced open-source AI framework that unifies video understanding, text-to-video generation, image-to-video generation, in context video generation, and free form video editing under a single multimodal instruction paradigm.

It uses a dual-stream architecture combining a Multimodal Large Language Model (MLLM) for interpreting complex instructions (including visual prompts) with a Multimodal DiT (MMDiT) for high quality video synthesis.

The model excels at task composition, zero-shot generalization to unseen editing instructions (like green screening or material changes), and transferring editing skills from image data to video.

Image To Video Text To Video Video Editor Video Generator

Pricing

Pricing Model

Free

Starting Price

$0/Month

Key Features

Unified handling of video generation and editing via natural language instructions
Dual stream design with MLLM for instruction understanding and MMDiT for video diffusion
Text to video and image to video generation with high fidelity
In-context and free form video editing, including zero shot unseen tasks
Support for visual prompts, task composition, and generalization from image to video editing

Pros

Single model replaces multiple specialized tools for video tasks
Strong generalization to unseen editing instructions without explicit training
Open source with model weights, inference code, and GitHub repo available
State of the art or competitive performance across multiple video benchmarks
Enables creative workflows like combining editing with style transfer

Cons

Requires significant GPU resources for inference due to diffusion-based generation
Still in research phase with potential setup complexity for non-experts
No hosted demo or web app; requires local/self hosted running
Video length and resolution may be limited compared to commercial tools

UniVideo is excellent for AI researchers, developers, and advanced creators who want a powerful, unified open source solution for generating and editing videos through simple instructions, especially valuable for experimentation and custom pipelines.

FAQs

What is UniVideo?
UniVideo is an open source unified AI model that handles video understanding, generation (from text or images), and editing all within one framework using natural language instructions.
Is UniVideo free and open source?
Yes, it’s completely free and open source. Model weights and inference code are available on Hugging Face (KlingTeam/UniVideo) and GitHub (KwaiVGI/UniVideo).
When was UniVideo released?
The research paper was published on October 9, 2025 (arXiv:2510.08377), with model and code released shortly after for community use.
What makes UniVideo different from other video AI tools?
Its dual stream design (MLLM + MMDiT) allows true unification of tasks, strong zero shot editing generalization (e.g., green screening without training), task composition, and visual-prompt support, outperforming many task specific models in benchmarks.