Stable Video Diffusion
A video generation diffusion model released by Stability AI, employing a DiT or DiT-variant architecture. It places substantial demands on GPU memory and computational resources.
Core Facts
Timeline (last 90 days)
The fundamental design goal of AnimateDiff and SVD is generation rather than precise editing; they do not natively support controlled editing that locks the background and replaces only the foreground person, and require additional condition injection such as ControlNet, depth maps, or optical flow constraints to improve controllability
视频生成存在扩散和自回归两大主流技术路线，扩散路线以 Sora、Stable Video Diffusion 为代表，自回归路线以 Google 的 VideoPoet 和 Meta 的部分工作为代表
Stable Video Diffusion 的训练流程分为三个阶段：低分辨率图像预训练、精选视频数据视频微调、高质量数据高分辨率微调
Stable Video Diffusion的训练流程分为三个阶段：低分辨率图像预训练、视频微调、高分辨率微调
SVD（Stable Video Diffusion）探索了在潜在空间中进行时间插值的方法以降低计算开销
All Facts (10)
Starting in 2023, diffusion models such as Stable Video Diffusion and Runway Gen-2 began to dominate the video generation market.
70%UnverifiedDiffusers集成了Stable Video Diffusion等多种视频生成管线，加入了时间注意力机制处理时间一致性问题
90%Unverified2023年后，以Stable Video Diffusion、Sora为代表的时序扩散模型开始将视频理解为时空连续体，能够在潜空间中同时建模多帧之间的运动关系
85%UnverifiedState-of-the-art video generation models such as Sora and Stable Video Diffusion both adopt DiT or its variants
70%UnverifiedThe underlying technology of image-to-video is mostly based on diffusion models, with representative products including Runway's Gen-3 Alpha, Pika Labs, and Stability AI's Stable Video Diffusion.
60%UnverifiedThe fundamental design goal of AnimateDiff and SVD is generation rather than precise editing; they do not natively support controlled editing that locks the background and replaces only the foreground person, and require additional condition injection such as ControlNet, depth maps, or optical flow constraints to improve controllability
50%Unverified视频生成存在扩散和自回归两大主流技术路线，扩散路线以 Sora、Stable Video Diffusion 为代表，自回归路线以 Google 的 VideoPoet 和 Meta 的部分工作为代表
50%UnverifiedStable Video Diffusion 的训练流程分为三个阶段：低分辨率图像预训练、精选视频数据视频微调、高质量数据高分辨率微调
50%UnverifiedStable Video Diffusion的训练流程分为三个阶段：低分辨率图像预训练、视频微调、高分辨率微调
50%UnverifiedSVD（Stable Video Diffusion）探索了在潜在空间中进行时间插值的方法以降低计算开销
50%