This week, Openai released its latest AI video generation model, the Sora 2, touted it as a “big leap” of space. When Sora hits the public, they need to compete for market share in busy markets, including major competitors who are rapidly gaining Steam. This is bytedance of a Chinese company that owns Tiktok.
In the past few months, ByteDance has released Seedance, an AI video generator that many users have already called the best in the world, and a new version of the elite image model, Seedream. According to the website analytics site AICPB.com, LLM's Doubao has 150 million active users.
Bytedance's AI advances are a prime example of how Chinese AI companies are quickly catching up to American companies despite chip export controls. Their models are high quality and inexpensive, attracting consumers all over the world, including in the US. However, while these models have attracted many users, there are many concerns that plague many of the cutting edge models. They can create affordable deepfakes that are indistinguishable from reality.
Reach the frontier
Over the past year, Bytedance has gathered top AI talent, hired former vice president of Google Deepmind, led basic research into AI, and separated other engineers and researchers from Alibaba and other startups. Financial Times It was reported in December. It also invests billions of dollars in infrastructure, including advanced Nvidia chips.
Bytedance released the first iteration of video model Seedance in June and the new image generator Seedream 4.0 in September. The model is accessible in the US via third-party platforms.
Jobin Jonny, a designer based in Kerala, India, first discovered Seedream in late August and was particularly impressed by his imagination of someone in his area. “The generated face carried the exact characteristics and details of a real Kerala man,” he says.
Johnny says that Sitance is now his favorite video model, especially when it comes to how he sees physics and the movements of nature. It doesn't hurt that sea dance is so cheap. The third-party platform Freepik costs half the credit of Google's VEO 3. Social media encourages several AI influencers to switch to Bytedance's products based on their price range.
Tiezhen Wang, an engineer at the machine learning platform, tested the tool as he hugged his face. He designed a “amazing” poster for Sea Dream, and says that Seedran “glows in maintaining image-to-video tasks, styles and character consistency.
Eric Lu, co-founder of online video editing software program Kapwing, has been providing AI image generation to its customers for several years, starting with stable spread. When Seedance and Seedream came out, his team conducted internal tests to compare fast adherence, image quality, speed and cost with American competitors. “And it wasn't close. The model is excellent in every way,” he says.
Lu quickly switched Kapwing's default AI image model to seed run and sea dream, away from the American model. “It was pretty easy because we not only save money, but also give users quality output,” he says.
“Unlimited” AI
However, this quality increase has many implications. First, it shows that Chinese companies have successfully navigated US chip export controls, designed to slow them down. information In December, bytedance reported that by renting outside of China, it has access to advanced Nvidia chips. The company is rapidly expanding its data center use in Malaysia.
Additionally, the price range of bytedance's tools allows many new users to rely on AI to create realistic images. The affordability and accessibility of these tools can reverse workflows in the advertising, marketing and stock footage industries. “Why buy a clip when you can generate the shots you need right away?” One X user wrote about the seed run in a thread.
As these surreal AI tools expand, deepfake threats and misinformation grow. In June, it was discovered that Google's video model Veo 3 produced realistic clips containing misleading or inflammatory information about a news event. After contacting Google about these videos, the company said it would start adding visible watermarks to the videos generated by the tool.
read more: Google's new AI tool generates persuasive deepfakes of riots, conflicts and election fraud
When I tested many of the same prompts used in Google's model using Seedance via Capcut's Dreamina tool, I actually rejected many of them because they violated community guidelines. Still, the model, like its competitors, has created decent realistic footage that could be shared as misinformation on social media, such as this video, created through the sowing of US soldiers providing assistance to Palestinian refugees. Representatives of bytedance did not respond to requests for comment.
The realism of the bytedance model also raises questions about copyright and portrait issues. Chinese scholars argue that China takes a “moderate generosity” regulatory approach in terms of training model on copyrighted materials. This will be displayed in the model output. For example, one X user posted a Sea Dream image of Joker from Heath Ledger, Harley Quinn from Margot Lobby, and Catwoman from Michelle Pfeiffer joining them in a dive bar. Another created images with Spider-Man, Batman and Superman.
Kapwing's Lu says Seedrun and Sea Dream are particularly willing to recreate copyrighted characters, whether they're Mickey Mouse or minions. “I think there's more scrutiny in America in that we source the content we're training for some of these big labs,” he says. “I think in China there is an unlimited ability for researchers to obtain the data they need and train them on it.”
Selina Xu, China and AI policy lead in Eric Schmidt's office, says it is “expected” to train models with user-generated video data from social media platforms. She adds that the video generation model is a “growthful revenue stream for AI companies.”
Time was able to create the image of “Young Brad Pitt and Leonardo DiCaprio shaking hands” through Capwing's Sea Dream. Some members of Congress, including Marsha Blackburn, are attempting to pass laws that protect the visual likeness of individuals and creators from digital replicas created without consent. However, such laws are still quite far from being passed.
Meanwhile, American companies have begun to pay attention to these Chinese AI giants, forcing them to publicly address them through copyright protection. In September, Disney, Warner Bros Discovery, and NBCuniversal sued Chinese company Minimax for copyright infringement of “Wilful and Brazen.”
“It's hot on fire”
American labs claim that because their Chinese counterparts have a loose attitude towards copyright, they should be able to train copyrighted materials, creating new images in ways that are transformative and protected under fair use. Earlier this year, Openai announced it would relax its content moderation rules, leading to a wave of internet-rich studio greeting memes.
read more: How those studio ghibli memes are signs of Openai's Trump era shift
“I'm not sure this is being driven by Chinese companies. The Open opened some locks to some extent in March,” says Maribeth Lau, an AI ethics researcher at the AI Accountability Lab at Trinity College, Dublin. She says that the ability of Baitedan models to create copyrighted characters and portraits of real people “unfortunately adds heat to the scramble fire at any cost, regardless of what kind of legal or ethical meaning.”
Rauh has many concerns about the spread of Deep Fark Tools, including that it could lead to increased harassment and misinformation and threaten the data privacy of users. “People have a very obvious interaction with these models: whether they're interested in generating, how to adjust them, or whether they're putting images of portraits of real people,” she says. “That's all the data that's at risk.”
Katharine TrendaCosta, director of policy and advocacy at the Electronic Frontier Foundation, argues that education is key to mitigate the risk of deep falk. “We've reached this strange point where we can produce everything, but no one believes anything anymore,” she says. “But we never solve the underlying issues. We continue to target new technologies, not media literacy or ways to teach or evaluate sources of analytical skills.”
