Stability AI Makes a Comeback: Stunning New Video Generation Demo
-
Is Stable Diffusion making a triumphant return?
Stability AI CEO Emad Mostaque's latest tweet with four videos has sparked endless speculation.
Many netizens suspect this is a demo of a new version of Stable Video Diffusion.
Judging by the results, the clarity, consistency, and smoothness of the visuals are truly astonishing. For example, this plush bear is diving with swimming goggles.
The plush bear moves very smoothly, even blinking its eyes. There are also rich details of seawater in the background.
Additionally, Emad Mostaque himself has been absent from social media for several weeks since January 1st of this year, never appearing online.
Netizens feel that this seems more like holding back for a big announcement. Stability AI is back? (Once thought Stability AI was falling behind in the competition)
Judging from the released 4-second demo, this appears to be an upgrade to the SVD video generation function.
However, it's still unclear whether it's based on text or image generation. Let's directly experience the effect.
There's also this anime-style night street scene. Last November, Stable Video Diffusion (SVD) was officially released. At that time, in addition to supporting image and text-to-video generation, it also supported 3D object synthesis.
According to the plans revealed at the time, SVD was just the foundational model, with further expansions planned to build out the entire Stable Diffusion ecosystem.
Feedback from users on the initial demo included issues with incorrect lighting and overall incoherence (flickering between video frames).
Comparing the current demo to a similar video example: a green robot lying on a bed. It can be seen that the details of the robot, as well as the nearby cup lamp and even the folds of the bed sheets and pillows, have become more abundant.
Moreover, there are obvious light and shadow reflections on the robot's face.
However, more update information will only be known after the official version is released.
It is worth mentioning that after this tweet, Emad Mostaque posted a similar one. No text, just four static images.
Netizens have expressed their anticipation. "Hope this model can surpass SDXL in all aspects and iterate better in a timely manner." "Is this SD3?"
But what on earth is this globe tomato???
On the same day, they released StableLM2 1.6B, a small language model with only 1.6 billion parameters. It has been trained in multiple languages including English, Spanish, German, French, Portuguese, and Dutch. A few days ago, they just released their first major language model of the year, Stable Code 3B.
With such frequent progress, is Stability AI really making a comeback?
The reason we talk about a comeback is that Stability AI was indeed in the spotlight during the last wave of text-to-image models. The explosive popularity of Stable Diffusion once made it the most talked-about AI unicorn globally. However, with the emergence of Midjourney and frequent progress by OpenAI and Google, coupled with internal instability—including allegations of the CEO plagiarizing achievements, misusing funds, unclear business models, and delayed salary payments—Stability AI once found itself embroiled in controversy.
Last November, the company announced its entry into the video generation sector, at a time when popular players like Runway and Pika had already emerged.
Now, with the suspected unveiling of a new version of SVD, the attention it has garnered is not only due to the company itself but also closely tied to the currently booming video generation sector. Many netizens bluntly stated: It's one wave of competition after another, with each company trying to outdo the last.
Whether it's comparing version releases and iterations—recently, domestic giants like Alibaba and ByteDance have frequently announced new progress—or focusing on generation effects, clarity, and smoothness. Not long ago, Moonvalley emerged out of nowhere, astonishing many.
Or competing in areas like controllability...
Just a few days ago, Runway's Gen-2 evolved once again— With just five brushes, by swiping across an image, the originally still birds start moving individually.
So, in the field of video generation, are you optimistic about Stability AI?