Snap Video

Snap VIDEo is an open research project from Snap Inc.’s AI research team, designed to advance the frontier of generative video modeling. Unlike many Text-to-Video tools built on image diffusion cascades or adapted image models, Snap Video adopts a native video-first APProach—treating time as a first-class dimension from the ground up. This enables more natural motion modeling, stronger temporal consistency, and richer dynamic detAIl without relying on post-hoc frame interpolation or temporal stitching.

What it does

Snap Video transfORMs plain English text Prompts into short, high-resolution video CLIPs. It accepts descriptive inputs—such as 'a Golden Retriever chASIng a red ball across a sunlit lawn'—and synthesizes corresponding video sequences with coherent object motion, lighting continuity, and spatial plausibility. The system is not intended for commercial production but serves as a technical benchmark and architectural blueprint for next-generation Video Generation.

Key capabilities

• Video-prior modeling: architecture optimized specifically for video data, not repurposed from image models—enabling better handling of motion redundancy and temporal dependencies.
• Extended EDM Framework: An evolution of the Elucidated Diffusion Model tailored for spatIoTemporal tensors, improving stability and fidelity across both space and time dimensions.
• FIT Transformer backbone: A flexible, sequence-efficient vision-language transformer that captures long-range temporal relationships while maintaining computational trACTability.
• Deterministic Sampling & classifier-free guidance: Enables precise control over text-video alignment during inference without stochastic artifacts.
• Joint space-time modeling: Learns motion patterns holistically rather than per-frame, resulting in smoother transitions and physically plausible dynamics.

Who it is for

Snap Video targets AI Researchers, academic developers, and advanced practitioners exploring foundational video generation techniques. It is especially valuable for those studying diffusion-based spatiotemporal modeling, efficient transformer design for video, or scalable training strategies for large-scale multimodal generation. End users seeking plug-and-play video creation should note this is a non-commercial research release—not a SaaS product.

How to get started

The model weights, training code, and Inference scripts are publicly available on GitHub under an open research license. Users need moderate GPU resources (e.g., multiple A100s) to fine-tune or run inference at scale. documentation includes configuration guides, Prompt Engineering tips, and evaluation metrics. No web interface is provided; usage requires local setup and command-line execution. Researchers may adapt components for domain-specific tasks like video editing or conditional motion transfer.

Pricing and access caveat

Snap Video is Freely accessible as open-source research software with no licensing fees. However, it is not a hosted service—there is no official web app, API, or cloud platform. deployment requires technical expertise in PyTorch, distributed training, and video preprocessing pipelines. Performance, resolution, and ouTPUt length may vary depending on hardware and implementation choices. commercial use is subject to Snap’s research license terms, which restrict redistribution and require attribution.