vita-epfl/Stable-Video-Infinity

What it solves

Stable Video Infinity (SVI) addresses the limitation of short video generation in AI models, enabling the creation of arbitrarily long videos (potentially infinite) while maintaining high temporal consistency, plausible scene transitions, and a controllable storyline without quality degradation or "forgetting" over time.

How it works

SVI uses a hybrid causality approach: it employs bidirectional attention within individual video clips to ensure high visual quality and then connects these clips sequentially using clip-by-clip causality. The system is implemented as LoRA adapters tuned on top of base models like Wan 2.1 and Wan 2.2, requiring minimal training data to achieve long-range generation capabilities.

Who it’s for

It is designed for creators, filmmakers, and animators who need to produce long-form AI video content, such as short films, single-scene animations, or storytelling videos, where consistency across many minutes of footage is critical.

Highlights

• Infinite Length: No inherent limit on duration, capable of producing videos lasting 10 minutes or more.
• High Consistency: Maintains visual stability and prevents color drift or content degradation across long sequences.
• Versatile Generation: Supports multi-scene films, single-scene animations, and conditioned generation (skeleton or audio).
• Efficient Training: Uses LoRA adapters, making it easy for users to create their own SVI variants with very little data.
• Open Source: Provides open training scripts, evaluation tools, and datasets.