UniEdit

UniEdit is an open research Framework that reDeFines how video editing can be perfORMed without model customization. Developed collaboratively by researchers from Zhejiang University, Microsoft Research ASIa, and Peking university, it leveRAGes the power of large pretrAIned Text-to-Video diffusion models to enable precise, instruction-based edits—directly from natural language Prompts. Unlike traditional Video Editing tools requiring manual keyframing or labor-intensive AI fine-tuning, UniEdit operates entirely out-of-the-box: users upload a source vIDEo and provide simple text instructions, and the system delivers edited ouTPUt while preserving spatial structure, temporal coherence, and visual fidelity.

What it does

UniEdit performs unified video editing across four core dimensions: motion, APPearance, background, and object identity—all guided solely by text. It does not train or adapt any underlying model. Instead, it uses inveRSIon techniques to map the input video into a latent noise representation, then steers the denoising process of a fRozen UNet backbone using cross-attention conditioned on user-provided text prompts. Crucially, it introduces aUXiliary branches—spatial self-attention for content preservation and temporal self-attention for motion guidance—to ensure consistency where editing isn’t intended.

Key capabilities

• Motion Editing: Replace ACTions in context—e.g., changing 'playing guitar' to 'waving hand'—while retaining subject identity and scene GEOmetry.
• style transfer: Apply artistic transformations like 'oil painting', 'cyberpunk', or 'watercolor' to entire video sequences without artifacts.
• background replacement: Seamlessly swap environments—e.g., moving a person from a studio to a mountain vista—while maintaining lighting and perspective alignment.
• Object Replacement: Substitute both rigid (e.g., a chair) and non-rigid (e.g., a waving flag) objects, respecting occlusion and motion dynamics.
• No training Required: All Operations run inference-only; no per-video tuning, LoRA adaptation, or dataset preparation is needed.

Who it is for

UniEdit targets AI researchers, creative technologists, and early-adopter filmmakers who need rapid prototyping of video edits without ML engineering overhead. It’s especially valuable for academic exploration of diffusion-based spatIoTemporal editing, as well as for designers Testing visual nARRatives before comMITting to full production pipelines.

How to get started

The framework is accessible via its official website and GitHub repository (source code availability may vary). Users upload a short video—typically under 3 SEConds—and enter a descriptive prompt such as 'make the Dancer perform ballet in a neon-lit city'. The system processes the request server-side and returns the edited CLIP. No local GPU or installation is required for basic use, though advanced users may deploy the code locally for customization.

Pricing and access caveat

UniEdit is released as open-source research software with no Commercial License restrictions. It is free to use, modify, and extend. However, official hosted web dEMOs may have usage limits or queue delays. Since it relies on large diffusion models, Inference speed and resolution depend on backend infrastructure—users should expect variable latency and output quality based on input length and prompt specificity. Versioning, official support channels, and long-term maintenance are not guaranteed, as the project remains research-oriented.