Almost a year has passed,OpenAI announced a 12-day live stream and for the first two days, everyone was staying up late complaining.
On the third day, Ultraman finally released Sora, and the shocking news was that the subscription for unlimited generation in 1080p pro mode costs up to $200!
Everyone grumbled and opened their computers, ready to subscribe.
As soon as the website was released,Sora’s traffic surged, forcing registration to stop on the morning of the 10th. Luckily, we managed to get in just before it closed. After a night of payment processing, we finally reached the successful subscription page!
Almost 1500 Chinese Yuan, what kind of results can we expect…
Actually, after nearly a year of explosive development in DIT technology, people’s expectations for Sora aren’t that high anymore, especially since there are many domestic video models launched like KeLing, JiMeng, HaiLuo, Vidu, PixVerse, HunYuan, ZhiPu, etc., with good results from various closed-source and open-source DIT architectures. People are already used to the generation of real AI images.
However, Sora, as the first DIT to release a Demo, still holds a special place in the hearts of AI enthusiasts, and everyone is biting the bullet to place an order. (Honestly, the cumbersome payment process can deter 95% of people; domestic Visa cards are not accepted, and we finally found a beautiful Visa card from the U.S. to complete the payment, thanks to zs classmate…)
Now, let’s get to the main topic!
Website:https://sora.com/
Everyone can first watch our Sora test comparison video and the experimental short film made with Sora!
—-
1. Basic Function Settings
Style presets, aspect ratio, resolution, duration, concurrency, text-to-video, image-to-video, etc.
Once you enter the webpage, you can intuitively see the UI for generating content, includingaspect ratio, resolution, duration, concurrency, which is very intuitive. The UI design is quite good, but why can’t you just drag and drop images to the webpage for uploading? You have to open a file selection window, which is a bit cumbersome…
Currently, it supports 3 aspect ratios, 3 resolutions, 4 durations, and 3 concurrency levels.
Next to the image upload button, there is a style preset, which is essentially a style filter pre-set for text-to-video, allowing for a unified generation style. However, this may not be very meaningful for professional users, but having style customization is a good effort to add some highlights.
Text-to-video and image-to-video functions don’t need much introduction; just write the text prompt and upload the image, and you’re good to go…
Sora Text-to-Video/Image-to-Video Effects:
The following videos are sourced from Vicky
Sora’s text-to-video still has many issues, such as semantic understanding, understanding of physical space, and understanding of direction, often resulting in incorrect movement directions for the main subject.
· Multiple limbs appearing is still a frequent occurrence.
· Unnatural hand movements, character twitching, overly strong 3D effects, poor anime quality, etc. Text-to-video is Sora’s main feature, and if it’s this abstract, then image generation is even worse.
· The quality of image generation is so poor that it cannot be described in natural language; only the results that can be edited into the video are relatively better.
· Image generation quality mainly suffers in several areas: first, it cannot recognize both foreground and background simultaneously, with poor understanding of physical space; the background often remains static; second, there are frequent and random cuts, as creators will edit, theoretically we don’t need cutting functionality, we just need good movement within a single shot; Sora’s cuts not only change the composition but also directly alter the style, which is very frustrating; third, there are issues with movement range and generalization; the poor movement range of Sora’s image generation makes me not want to complain anymore, and it often changes color saturation and has flickering issues.
I also made a comparison video of Sora’s text-to-image and image-to-video:
—-
2. Storyboard Function
The storyboard function allows for more precise control of camera movements within the same time segment using text, images, and videos. Previously, generating 5-10 seconds of footage could only be controlled by prompts, but this storyboard function allows you to specify the exact duration you want to control, including text, images, and videos,but currently does not support uploading videos with people.
The maximum duration can reach 20 seconds, and multiple layouts can be added; the idea is great, but the reality is…
Storyboard Function Effects:
It’s actually quite convenient and quick, usually sufficient for making concept videos and trailers. However, it often encounters semantic understanding issues. Sometimes when you input multiple prompts, like 4 segments, it often only recognizes one or two segments, which can be quite frustrating.
The image storyboard function is currently a disaster. Uploading images is really slow…
I even suspect that Sora’s image generation has nothing to do with DIT… how can it feel so much like a PowerPoint presentation?!
I basically don’t test text-to-video because it cannot serve as a productivity tool, but apart from text-to-video, Sora has nothing else to test.
Honestly, if it weren’t for Sora, and if I hadn’t already spent $200, I definitely wouldn’t spend so much time on this, trying to seek the value of that $200 and find Sora’s highlights…
—-
3. Video Editing Function
The video editing function is another highlight feature that Sora is promoting this time, but under the premise of such poor image generation model capabilities, these features seem useless to me…
Sora has put a lot of effort into researching how to create a product that serves creators well, designing features and interactions with great creativity, but with the current poor capabilities of the image generation model, how can we use it! How to use it!!! Oh dear!!!
Let’s still politely introduce the features…
The subscription has two tiers:$20 only generates 50 720p watermarked videos,$200 can generate 500 1080p videos + unlimited slow motion + watermark-free downloads.
4. Sora Generation Cases
Some artists have created Sora samples, and to achieve this, I really don’t know how many times they had to draw cards…
Sora’s realistic light and shadow performance is quite good, but this certainly isn’t something that can be achieved in just one or two draws; most shots look like text-to-video.
Conclusion:
1. Sora’s image-to-video is currently really hard to use.
Why is Sora’s image recognition so poor, why!!! I tested image-to-video and really reached the point of breakdown; it’s too hard to achieve good quality.
2. Sora’s text-to-video effects are acceptable but require drawing cards, making it difficult to create continuous narrative content.
Text generation struggles to maintain consistency between characters and scenes, so it can only lean towards consciousness flow and visually conceptual “art films” in text generation.
3. There are two highlights: first, the storyboard concept is quite good, but the model capabilities do not keep up; second, text-to-video has a strong sense of realism, but image-to-video is lacking, making it redundant.
Writing good text prompts for text generation and familiarizing yourself with product rules may still lead to some good works, even in the case of drawing cards to the point of madness…
People’s expectations for it were indeed too high before.
Currently, domestic AI video products are catching up quickly and are very competitive; people have already seen too many good results, and the era of casually producing 4-second videos is gone. Besides model capabilities, users also care about draw rates, interactivity, package cost-performance ratios, etc. From any angle, Sora currently does not have enough advantages.
4. If you just want to make AI videos, there’s no need to buy it; $200 is not worth it, and even $20 is not necessary. However, if there are Sora fans or rich folks who enjoy making text-to-video concept films, then let’s pretend I didn’t say anything…
Sora no longer brings surprises, and domestic products are closely following.
AI technology can iterate, innovate, and truly mature in the future, which is what every creator hopes to see.
But we are still in the early stages; which manufacturers can persist? What kind of technological ecosystem will lead to the end?
No one knows.
It’s still too early to draw conclusions. But we sincerely hope that day will come.
—-
A little Easter egg:
I hastily created a Sora AI experimental short film called “MOVIE ELEMENTS”
Can everyone guess what classic movie elements are in the short film?
Benefits
Address:https://www.yuque.com/frannnk7/aidesign?# 《AI Knowledge Base Collection PRO》
Historical Articles
Interested or curious friends are welcome to join the AI learning group.