Stability AI releases SVD 1.1 for more consistent AI video

Anyone who has experimented with AI image-to-video tools knows the disappointment of a promised cinematic shot turning into a frozen, distorted mess where the camera barely crawls.

Stability AI is trying to fix that stiffness. The company just released SVD 1.1, an upgraded version of its open-weight latent diffusion model, Stable Video Diffusion.

Built as a fine-tuned update to SVD-XT, SVD 1.1 focuses squarely on delivering more consistent motion and realistic frame transitions without requiring endless parameter tweaking.

Locking in the sweet spot for smoother motion

When Stability AI introduced SVD and SVD-XT in November 2023, the models showed promise but had noticeable limitations. The base version output 14 frames over four seconds, while SVD-XT bumped that to 25 frames.

In practice, however, early users regularly ran into issues. Outputs frequently lacked photorealism, character faces morphed unnaturally, and camera pans were often painfully slow or nonexistent.

With SVD 1.1, Stability fine-tuned the model by locking in specific default settings: a 6FPS conditioning rate and a motion bucket ID of 127. Basically, they dialed in the parameters that produce the most reliable motion out of the box, sparing creators from endless trial and error.

According to Stability AI CTO Tom Mason, the model generates 25-frame, four-second clips at 1024x576 resolution. The model weights are available to download on Hugging Face, free for research use. Commercial deployment requires a membership starting at $20 per month.

For developers building external apps, Stability also rolled out an API. The endpoint handles motion strength control and exports 24fps MP4 files with frame interpolation, supporting landscape, square, and vertical aspect ratios.

Why it matters

We think Stability AI's biggest win here is not about a massive architectural leap, but about predictability.

For developers and video creators, the primary roadblock with open-source video models has been the abysmal hit rate. Spending compute on generations that fail half the time kills production viability. By stabilizing the baseline motion, SVD 1.1 makes local and cloud-based video pipelines much more practical to deploy right now.

This release also highlights an interesting split in the generative video race. Rivals like Runway and Pika are focused on polished consumer web interfaces, rolling out features like motion brushes and localized element swaps. Yet neither currently offers an open API for third-party developers.

Stability AI, on the other hand, seems content playing the infrastructure layer. By releasing both open weights on Hugging Face and a production-ready API, it is positioning itself as the default engine for external software builders.

That said, four seconds of video with potential legacy glitches shows that latent diffusion still has a ceiling. Stability will need more than modest fine-tunes if it wants to stay ahead as the field rapidly scales up.