On the eve of the Lunar New Year, as brushstrokes bled into galloping ink horses and flower spirits appeared to breathe beneath the stage lights, China’s most-watched television event crossed a technological threshold.
In one performance, the poet Bai Juyi drifted across shimmering waters as he recited his verse.
In another segment, Ming dynasty painter Xu Wei splashed ink that dissolved into blooming petals.
Elsewhere in the program, the legendary beauty Wang Zhaojun turned back toward the Central Plains with a sweep of her pipa.
Some viewers described the separate sequences as stirring a renewed sense of connection to traditional Chinese culture.
Only after the segment ended did the hosts disclose that its visuals were generated using the image and video capabilities of Doubao, the large model developed by ByteDance’s Volcano Engine. For many viewers, that reveal was the twist: artificial intelligence had reached the gala stage in full cinematic form.
This year’s broadcast may very well go down as the most AI- and tech-intensive edition in the gala’s history,
A national AI spectacle
An ink-and-flower performance, whose title roughly translates to an ode to the flower deities, was one of the night’s most discussed segments.
Another program built around the theme of dreams featured dancer Liu Haocun stretching across the stage while five hyperreal digital avatars mirrored her movements behind her. As cameras shifted and lighting changed, the avatars’ perspective and shadows adjusted in real time, suggesting live rendering rather than pre-produced footage.
A comedy sketch starring veteran performer Cai Ming introduced a wisecracking robot grandson capable of backflips and punchlines. In one exchange, the robot joked about whether a grandmother would save her human grandson or her robot grandson first if both fell into a river. The humor drew additional resonance because audiences had recently watched Doubao spar with entrepreneur Luo Yonghao at a public event.
Throughout the evening, hosts repeatedly encouraged viewers to open the Doubao app to generate Lunar New Year greetings or AI-crafted profile avatars. The message was understated but clear: don’t just watch AI, use it.
On Lunar New Year’s eve alone, Doubao reported 1.9 billion AI interactions. According to figures cited during the broadcast, its holiday campaign generated more than 50 million festive avatars and over 100 million greetings. At peak load, when a host initiated a second round of interaction, Doubao’s model inference throughput reportedly reached 63.3 billion tokens per minute. These figures were provided by the company during the program and have not been independently verified.
Meeting the gala’s demands
The most consequential test of AI’s utility came long before the live broadcast, inside rehearsal rooms.
About a month before the show, the gala’s director team approached Volcano Engine with a request: animate a static ink painting in the style of modern master Xu Beihong, featuring several distinct horses. Even a simple trotting motion would suffice.
At the time, ByteDance was approximately 30% through training its next-generation flagship video model, Seedance 2.0. Engineers could not guarantee results.
Ink wash visuals posed a distinct technical challenge. Training data for traditional Chinese ink painting is limited, constraining model development. Many Western-developed models therefore struggle to replicate the aesthetic logic of Chinese freehand brushwork, which emphasizes expressive strokes over photorealistic detail. Without storyboards or motion references, animating this style raises a fundamental question: how should ink move?
The team said it adopted an iterative approach, prioritizing usability over perfection. Instead of generating full sequences directly from text prompts, engineers created keyframes and built motion around them. According to ByteDance’s team, they did not fine-tune a separate model specifically for the gala. Instead, they pushed Seedance 2.0 to its limits at each training stage.
At peak intensity, engineers produced dozens to more than 100 video iterations per week, a pace that would be unusual in traditional film production. Director feedback fed back into the training loop.
The final result transformed a static painting into a minute-long sequence with dispersion, convergence, close-ups, and character interaction. Each horse retained its color and posture before reassembling into a unified tableau.
Seedance 2.0’s native output currently supports up to 720P resolution at 24 frames per second. To meet the gala’s HDR and 8K broadcast requirements, the team developed a separate enhancement pipeline to upscale the footage.
Creative control also evolved. Early in the process, engineers proposed solutions to directors still testing AI’s limits. Once the model’s usability reached what team members described as roughly 80–90%, directors resumed full creative authority, issuing detailed artistic instructions that the model could execute with greater reliability.
A “Black Myth moment” for Chinese AI?
Across both the stage performances and the Doubao app’s interactive features, one name surfaced repeatedly: Seedance 2.0.
Its release ahead of the holiday drew attention within China’s technology sector. Game Science founder Feng Ji reportedly described it as a “Black Myth moment” for Chinese AI, referencing the breakout game Black Myth: Wukong.
What changed? Earlier video-generation models often operated unpredictably. Creators built detailed storyboards, generated multiple clips, and selected the few that were usable. Maintaining character consistency typically required substantial post-production.
By contrast, ByteDance said Seedance 2.0 can generate 15–30 second sequences with camera movement, consistent characters, and synchronized audio from relatively simple prompts, with a reported usability rate above 80%. Transitions between shots are designed to follow what’s described as a director’s intent, shifting the creative workflow from trial and error toward guided generation. These performance metrics are based on company disclosures.
According to the team, the improvement did not stem from a single breakthrough. Instead, it resulted from systematically addressing weak points in training and inference. Following the industry’s broader turn toward reinforcement learning, including advances from companies such as DeepSeek, ByteDance increased investment in foundational models while separating model development from product iteration.
Cost reductions have compounded the impact. If the rate of usable clips falls to roughly one in two attempts, production costs can decline to about RMB 1.4 (USD 0.2) per second, a reported 68% reduction. That shift has reportedly contributed to a rise in AI-generated short videos across platforms including Douyin, Bilibili, and Xiaohongshu.
Getting a billion people to try AI
In 2015, when the gala’s hosts told viewers to “shake” their phones, the resulting WeChat “red packet” campaign accelerated mobile payments adoption across China. The infrastructure was ready. Users needed only a prompt.
AI infrastructure remains more computationally demanding. A Volcano Engine executive estimated that a red packet request consumes less than one ten-thousandth of a TOPS, or tera operations per second, in CPU power. Generating AI-powered visuals can require roughly ten TOPS per request, millions of times more computational intensity.
Why pursue mass participation now?
In a 1998 interview with BusinessWeek, Apple co-founder Steve Jobs said:
“A lot of times, people don’t know what they want until you show it to them.”
In the three years since ChatGPT gained global attention, startups have experimented with different paths to demand. In the US, Anthropic expanded in enterprise markets with its coding-focused Claude tools, while other companies continued scaling model size.
In China’s consumer-driven internet market, model capability and user acquisition tend to advance together.
As of December, Doubao’s large model reportedly surpassed 50 trillion daily tokens in usage, representing more than tenfold year-on-year growth, according to company figures. Volcano Engine’s 2025 revenue exceeded RMB 20 billion (USD 2.8 billion), also based on company disclosures.
Consumer adoption can drive enterprise procurement. A positive experience with Doubao or ByteDance’s Dreamina may influence purchasing decisions for Volcano Engine’s AI cloud services. At the same time, enterprise requirements can accelerate infrastructure upgrades.
Advanced models are a prerequisite. The more difficult task is translating technical capability into everyday utility, convincing first-time users that multimodal AI is not an abstract competition over benchmarks but a practical tool.
On this year’s gala stage, that translation may already have begun.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Xi for 36Kr.