OpenAI's new video generation tool, Sora, officially launched today with advanced features like Remix, Re-cut, and Storyboard, promising significant improvements in workflow efficiency for creators. Learn about its pricing, performance, and the future potential of AI-driven video creation.

Early this morning, OpenAI Sora officially debuted.

The launch event continued the fast-paced style of a "short drama," lasting about 20 minutes, hosted by CEO Sam Altman, Sora lead Bill Peebles, and others.

OpenAI stated on the X platform that since February, they have been developing Sora Turbo, a significantly faster version of the model. Today, they have launched it as a standalone product, available to Plus and Pro users.

Interestingly, due to the overwhelming popularity of Sora, a large number of users flooded the experience website, causing it to crash temporarily and halt registrations and logins. The inadequate service also prompted Altman to repeatedly reassure users on the X platform:

"Due to demand exceeding expectations, we will have to intermittently close new user registrations, and the content generation speed will slow down for a while. We are working hard to resolve it!"

Here is the experience link: Sora.com

Sora's interface is revealed with six powerful features—no need to learn editing anymore?

Similar to the Midjourney web interface, Sora also has its own user interface, where users can organize and browse generated videos and view other users' prompts and selected videos.

In the "Library," users can save their favorite or useful prompts for future use. Saved prompts can be viewed or edited as needed, significantly improving efficiency for users who need to repeatedly create similar content.

In terms of workflow, Sora's editing features distinguish it from other competitors.

For example, with the Remix feature, users can edit videos using pure natural language prompts, and control the degree of changes to the generated content through simple "strength" options and sliders.

The Re-cut feature intelligently identifies the best scenes and supports extending the scene in any direction.

The Storyboard feature is similar to a video editor, allowing users to link multiple prompts together to generate a longer video, making it easy to handle complex multi-step scenes.

With the Loop and Blend features, users can create seamless looped videos and perfectly blend different segments. The Style presets function allows users to preset and adjust the style of the generated content.

In terms of technical specifications, Sora supports video generation from 5 to 20 seconds, compatible with popular aspect ratios like 1:1 and 9:16. Compared to earlier versions, the generation speed has significantly improved.

Additionally, there are some details to note.

Sora Turbo adopts a flexible point-based pricing strategy, where the points required vary based on resolution and duration. If you are already a ChatGPT Plus or Pro member, you can use it without any additional cost.

For example, generating a 480p, 5-second video requires 25 points, while generating a 480p, 20-second video requires 150 points.

Moreover, if you use features like Re-cut, Remix, Blend, or Loop, and the generated content exceeds 5 seconds, you will be charged extra points. The more you use, the more it costs. Don't exceed the time limit, as exceeding it also incurs additional costs.

For subscription users, the $20 ChatGPT Plus plan provides 50 priority video credits (1000 points), supporting up to 720p resolution and 5-second duration.

The $200 ChatGPT Pro plan offers up to 500 priority videos (10,000 points), supporting up to 1080p resolution, 20-second duration, 5 concurrent generations, and watermark-free output.

OpenAI is also considering developing different pricing models for various types of users, which will be launched in early next year.

By the way, Sora does not support ChatGPT Team, Enterprise, or Edu users, and is also not available to users under 18.

Currently, users can access Sora wherever ChatGPT is available, except in regions like the UK, Switzerland, and the EU.

Real-world tests of Sora reveal its biggest shortcomings, but some scenarios are on par with professional-level work.

Well-known blogger Marques Brownlee used Sora a week in advance and shared his experience on YouTube.

He pointed out that the product still has some limitations.

In terms of physical simulation, the model's understanding of object motion is not deep enough, often resulting in unnatural movements or sudden disappearing objects. Specifically, when processing objects with leg movements, the leg positions are often confused, making the motion look unnatural.

Additionally, some video results appear as slow-motion, while other parts play at normal speed, creating an easily noticeable "awkward" effect. In short, Sora still hasn't resolved its old problems and lacks an understanding of the laws of the physical world.

Furthermore, Sora has not resolved the text generation issue, resulting in frequent text confusion, although the editing style, text scrolling, and news anchor-style generation are particularly realistic.

However, Sora excels in many scenarios.

For example, Sora performs excellently with landscape shots, generating drone footage comparable to professional material. It also performs decently with cartoon and stop-motion animation styles.

In terms of performance, a 5-second 360p video can typically be generated in 20 seconds. However, for 1080p or complex prompts, the generation time may extend to several minutes. With the influx of users, the generation speed has slowed significantly.

Many users also experienced Sora firsthand. For example, user @bennash attempted to generate a video, but the rendering took 22 minutes without success, and the website even temporarily stopped registering logins.

Blogger @nickfloats commented that while some specific visual effects were not preserved when converting images to video, the overall conversion effect was "clear and satisfactory."

Can Sora become the next "golden goose"?

Sora's System Card also lists some important details.

OpenAI believes that Sora lays the foundation for models that can understand and simulate the real world and will be an important milestone toward achieving Artificial General Intelligence (AGI).

The official blog mentioned that Sora is a diffusion model that begins with a video that looks like static noise, progressively removing the noise and transforming it into the final video. By processing multiple frames simultaneously, the model successfully solves a problem: even if the target temporarily leaves the field of view, it ensures consistency in the video.

Like the GPT models, Sora uses the Transformer architecture.

Sora employs the annotation technology from DALL·E 3, which generates highly descriptive labels for visual training data. As a result, the model can more accurately generate video content based on user text instructions.

In addition to generating video from text prompts alone, Sora can also generate video from existing static images, accurately animating the image content and preserving details.

To ensure the safe deployment of Sora, OpenAI has strengthened security measures based on its experience with DALL·E in ChatGPT and API deployments, as well as safety protections from other OpenAI products like ChatGPT.

• It is prohibited to use others' likenesses without permission and to depict real minors;

• Creating illegal content or infringing on intellectual property is prohibited;

• Generating harmful content, such as non-consensual intimate images, content for bullying, harassment, or defamation, or content meant to spread violence, hatred, or cause others pain, is prohibited;

• Creating and distributing content for fraud, scams, or misleading others is also prohibited.

All Sora-generated videos contain C2PA metadata, which can identify the source of the video, enhancing transparency and enabling source verification.

Unlike Flux, which gained attention for using real-life likenesses, Sora has set strict review standards for content uploaded that includes people. This feature is currently available only to a small group of early testers as a pilot function.

Over half a year ago, the initial trial of Sora received widespread acclaim on the internet.

However, if a year ago, people were still yelling "reality doesn’t exist" in front of demo displays, after a wave of various video models at home and abroad, our expectations have evolved to a point where it’s hard to be impressed by the same kind of products.

This change in attitude stems from a simple fact: as AI evolves from "barely usable" to "highly useful," user expectations also shift from "can it be done?" to "how well can it be done?"

Fortunately, Sora has not stood still amid the applause. Through deep collaboration with artists, significant improvements have been made in the workflow domain. Features like Re-cut, Remix, and Storyboard are quite practical.

The existence of both parties (client and provider) ensures that communication within the workflow is always essential. What AI can do is make this communication more efficient. The value of Sora is not in what it can do but in how it allows creators to step away from technical details and truly return to the essence of creativity.

At the same time, the $200 ChatGPT Pro subscription plan, which sparked heated discussions last week, now has a more reasonable price anchor. This plan also provides unlimited access to Sora. This product synergy is expected to unlock application scenarios and commercial value far beyond expectations.

Looking at the present, users' spending is never fake.

Qingling AI has delivered impressive results with a monthly revenue of tens of millions, and the potential of this blue ocean market is already becoming apparent. For OpenAI, which is still in the "burning money" phase, Sora is expected to become another golden goose after ChatGPT.

When Sora evolves from "usable" to "excellent" to "ingenious," perhaps one day in the future, we will realize that the real thing that doesn't exist is not reality but the limits of human creativity.