Sora's New Video Exclusively Released on TikTok: OpenAI Gains 100K Followers in 4 Days
-
Sora's new video has become a 'TikTok exclusive.'
Quietly, OpenAI officially entered TikTok, with addictive background music that makes viewers scroll endlessly, rapidly attracting followers:
In just 4 days, it gained 100K followers and 500K likes—and this was achieved without any promotion or advertising. How can human creators compete after this?
On OpenAI's newly verified TikTok account, a series of newly generated videos have been quietly updated.
Each one is distinctive and has sparked widespread discussion.
Take a look: a16z partner exclaimed that if this appeared in a news feed, one absolutely couldn't tell it was fake.
I've seen many AI videos, but this one feels like a turning point.
In real short video consumption scenarios, Sora's generational gap with competitors becomes even more apparent: so realistic that it requires special disclaimers stating it's not real. OpenAI is extremely concerned about the potential negative impacts of AI-generated videos being taken as real, going to great lengths to remind people 'This is fake! This is fake!' in various ways.
Netizens have started issuing warnings like this:
From now on, everything you see online is questionable. Don't believe your eyes.
There's an old saying that goes, 'On the internet, nobody knows you're a dog.' Now officially upgraded to the "Sora era, where no one knows you're a non-existent dog."
Below are a few "TikTok-exclusive" Sora videos to enjoy:
Ultra-macro tracking of an ant colony, with effects comparable to a documentary.
Even static images previously created by DALL·E 3 can be repurposed; with some Sora processing and electronic music, they transform into viral dance videos. They even released a 1-minute 24-second video specifically to demonstrate AI-powered video editing capabilities similar to TikTok effects.
This has left many short video creators feeling overwhelmed, with some considering quitting the platform.
Even the platforms themselves are facing unprecedented challenges from this technological disruption.
A showdown seems to be brewing between ByteDance, the world's highest-valued unicorn, and OpenAI, which recently jumped to the third position. The ultimate battle between the UGC king and AIGC newcomer is inevitable.
In fact, even before Sora's overnight success, ByteDance already possessed AI video technology internally.
One of them, Boximator, was even rumored to be the "Chinese version of Sora". The Boximator paper was published before Sora's release, led by ByteDance AI Lab Director Li Hang.
Actually, Boximator's functionality does not generate videos from pure text, but rather animates static images according to text prompts.
It builds upon ByteDance's PixelDance base model launched last November, incorporating new motion setting methods—— By selecting objects from images and setting motion endpoints or directly drawing trajectories, precise control over dynamic effects can be achieved.
In terms of implementation, Boximator functions as a plugin that works with existing video diffusion models. It achieves this functionality by freezing the weights of the base model and only training the control module.
For example, by drawing bounding boxes and trajectories, you can make a little Hachimi (pet) chase a ball along a specific path. In terms of results, although there are some bugs (the floating umbrella lacks a handle), Boximator's animation is significantly more complete and closer to the text prompt when compared to Pika 1.0 (middle) and Gen-2 (right).
From the perspective of material and physical effects, Boximator also performs better.
Technically, in terms of architecture, Boximator adds a new self-attention layer in each spatial attention block of the video diffusion model to process control tokens encoded with object IDs, hard/soft flags, and bounding box coordinates. To simplify the challenge of learning 'box-object' associations, Boximator also adopts a technique called self-tracking during the training process.
So, can Boximator be considered 'ByteDance's version of Sora'?
ByteDance responded to these rumors with a denial: Boximator is a technical research project in the field of video generation focusing on object motion control. Currently, it cannot be implemented as a mature product and still has a significant gap compared to leading international video generation models in terms of image quality, fidelity, and video duration.
Boximator builds upon previous research by the ByteDance team, including PixelDance (basic research) and MagicAnimate (focusing on human motion), as their latest achievement.
Overall, these technologies were not specifically developed for Sora, but in the context of the intense competition in AIGC, they still serve as a form of "strategic preparation". Additionally, earlier this year, Zhang Nan, the former CEO of Douyin, stepped down to focus on Jianying, riding the wave of AIGC.
It can be said that although the release of Sora was sudden, ByteDance was not unprepared. In fact, it had already sensed the technological trend in its business operations.
Compared to the inevitable showdown between OpenAI and ByteDance, other players in the upstream and downstream of the video industry seem to have been hit by an unexpected blow.
Sora's first strike actually targets Adobe. Five days have passed, and the stock price continues to fall, with market capitalization evaporating over $10 billion.
At this critical moment, Adobe has just urgently announced the formation of a 50-person AI research team.
However, the scale of this investment fails to demonstrate sincerity.
Out of 29,945 employees, 50 people account for only 0.16%. Adobe represents video production and editing tools that have been in use for nearly 30 years, serving content creators.
Without transformation, it risks becoming the next Kodak or Nokia, but the transition is also seen as highly challenging.
For instance, some point out that while Adobe possesses vast amounts of data, using past creators' works to train AI could strain relationships with these users.
This is precisely why Adobe's AI image-generation tool, Firefly, was not trained on data from its own creative community, Behance, leaving it somewhat constrained. For instance, judging from Adobe's past performance in AI-generated images, there was already a significant gap compared to Midjourney. Now, how will they address the technological disparity with OpenAI in the realm of AI video?
Apart from Adobe, the second most impacted are the stock material suppliers, with companies like Shutterstock also struggling.
Even Elon Musk agrees with the view that 'this industry might vanish.' However, Shutterstock has already secured a fallback by signing a cooperation agreement with OpenAI early on, transitioning from a content provider to a training data supplier.
Among the Sora videos that have been trending recently, many feature Shutterstock images.
Some have jokingly speculated that the four letters in Sora might stand for "Shutterstock Original, Remixed by AI."
References:
[1]https://www.tiktok.com/@openai
[2]https://boximator.github.io/
[3]https://twitter.com/bilawalsidhu/status/1760076742679552273