The Truth Behind Seedance Traffic Jam

“Currently, when creating AI videos, it’s either about generating content or waiting in line to generate it.” This self-deprecating remark is common among AI video creators today.

On social platforms like Douyin and Xiaohongshu, many creators are venting their frustrations under the topic of “Seedance Queue.” Some report that a one-minute generation task submitted at 9 AM remains in the queue by the time they finish work in the evening, with one creator noting, “After inputting the prompt, there are still 80,000 people ahead of me in line.”

Since its launch on February 12, ByteDance’s Seedance 2.0 video generation model has become a standard tool for nearly all AI short drama and short film teams due to its powerful generation and adaptation capabilities.

Peng Yuhong, one of the earliest screenwriters to enter the short drama field, brought her team on board with the launch of Seedance 2.0. However, they quickly found their efficiency hampered by long wait times.

“The servers are particularly crowded during the day; it’s only after 8 PM, especially during the early morning hours, that things run a bit smoother,” Peng told a reporter. Consequently, most teams in the industry have opted for off-peak production, even adjusting their working hours to night shifts.

Compared to traditional film production, AI video generation has a crushing cost advantage. Liu Shuai, CEO of Beijing Xunzhi Zhonghe Technology, shared that his company produced a popular 7-minute AI short film titled “Dilemma” with just him and the director working during the Spring Festival, consuming only a few thousand yuan in token costs.

In traditional filmmaking, excluding the fees for well-known actors, basic costs for the crew, location, and equipment can easily exceed 200,000 yuan. “Even if the waste rate for AI-generated videos is high, requiring repeated iterations, the costs are negligible compared to traditional filming.”

The influx of creators has led to a significant computational power gap, causing Seedance to become congested. However, recent reports indicate that the situation is improving, though not due to an increase in computational power.

An insider close to the Volcano Engine revealed that Seedance is redistributing computational power weights and intentionally “dumbing down” the service. “By reducing the computational allocation and model running precision for individual tasks, more users can be online simultaneously, supporting higher concurrency—at the cost of lower precision for individual tasks.”

The explosive demand and business opportunities surrounding Seedance have led ByteDance to quietly raise the usage thresholds.

Currently, the Volcano Engine has announced that the Seedance 2.0 API pricing is approximately 28 yuan per million tokens (with video input) and 46 yuan per million tokens (without video input). The cost to generate a 15-second video is about 15 yuan, equating to 1 yuan per second.

This price is significantly higher than the C-end “premium membership” rate of about 0.2 yuan per second. However, insiders indicate that the API interface allows access to a “full version” of Seedance without queues and with relaxed reviews. Yet, the Volcano Engine’s official website states that the API is currently only available to select commercial partners and is not fully open.

The aforementioned insider indicated that the API whitelist is primarily open to large film companies, content production firms, and specific institutions, with varying discounts for different organizations. Some institutions face a “minimum consumption” requirement of up to 10 million yuan per year. Most AI short drama companies do not reach this spending level and must resort to “purchasing packages” to access the service, giving rise to a new “broker” business in the market.

Additionally, it is reported that Jimeng is about to launch an AI comic production tool that will also integrate the Seedance model to capture the AI comic market.

15-Second Videos Can Take 8 Hours

Currently, the generation efficiency on Seedance varies drastically between day and night.

Peng Yuhong noted that generating a 15-second video during the day typically involves a wait of several hours, with extreme cases requiring over half a day. In contrast, during the early morning hours, the same 15-second content can yield results in two to three hours, or even faster.

To tackle the queue problem, Liu Shuai has completely restructured the production process. His team operates in two shifts: from midnight to 10 AM, they focus on video content production during the server’s lowest load period. By 9 AM, post-production colleagues arrive to edit the content generated the previous night. In the afternoon, the team uses self-developed tools to supplement shots and generate images for the night’s video production.

To maximize time utilization, many companies in the industry register multiple accounts to submit generation tasks simultaneously, while some teams work around the clock in shifts to seize computational power windows.

The core issue causing long queues at Seedance is the insatiable demand for computational power. “Currently, all major companies’ computational power is insufficient,” Liu Shuai stated.

AI video generation is inherently resource-intensive, requiring complex computations for image generation, motion continuity, light and shadow matching, and scene consistency, consuming far more computational power than AI-generated images. A 15-second, 1080P video requires approximately 300,000 tokens, directly limiting the server’s processing capabilities.

After Seedance’s surge in popularity, the user base has exploded, with not only AI short drama teams and film production companies entering in droves but also a vast number of C-end users and self-media creators, leading to a surge in concurrent requests and immense pressure on the servers.

However, recent improvements have been noted, with the wait time for generating a 15-second video no longer requiring “at least 8 hours.” At the same time, the success rate of video generation has fluctuated.

“Previously, generating two or three videos would yield one usable piece of material; now it often takes generating seven or eight to find one. The review process has also become increasingly stringent; even without real human materials and no pornographic, bloody, or violent content, using the same set of prompts, what was once fine yesterday might be flagged as a violation today,” one industry professional told reporters.

As mentioned earlier, Seedance is implementing a “dumbing down” of some individual tasks. To achieve high concurrency and access to a private API interface, users must meet the “minimum consumption” standards.

An insider close to the Volcano Engine noted that some users on the API whitelist have enjoyed token discounts but do not utilize all their computational power, leading them to resell access to other companies. “Recently, many scammers have emerged in the market, claiming they can help create ‘packages’ for 100,000 yuan a month, only to run off with the package fees.”

Additionally, it is reported that Seedance is also exploring overseas markets, with some individuals selling exclusive usage rights for as much as $2 million per year abroad.

AI Video Creation Is Not That Simple

The long queues are just the tip of the iceberg in AI video generation. While many believe that “AI has lowered the threshold for film production to the floor,” the actual creative process is far more complex and arduous than the public imagines.

Currently, the domestic AI-generated video sector has formed a competitive landscape with multiple tools, each with distinct advantages and disadvantages.

Liu Shuai stated that, overall, Seedance remains the clear leader in the domestic market, while Kuaishou’s Keling competes in certain scenarios.

In terms of overseas models, before the official launch of Seedance 2.0, Liu Shuai’s team frequently used Google Veo 3, which performed well in quality but struggled with content instruction control, making it challenging to meet creators’ refined needs. In contrast, Nano Banana excels in image generation and editing, suitable for fine-tuning storyboard images.

Peng Yuhong’s team is also using Jimeng and Xiaoyunque, both of which belong to the same technical system with Seedance as their underlying model. Jimeng’s video generation costs about 1 yuan per second, while Xiaoyunque has shorter wait times but higher costs, at about 2 yuan per second.

Both Liu Shuai and Peng Yuhong agree that there is no perfect tool for professional production teams; instead, they must combine multiple tools based on different production stages.

However, beyond the queues, the more challenging issues are the fine-tuning and consistency problems in AI video. The public often thinks that simply inputting a prompt will yield a perfect video, but in reality, creators spend 80% of their time battling AI’s “uncooperative” nature.

Liu Shuai recalled a particularly memorable case: the team needed a shot of an “actor lowering their head in contemplation and then slowly raising it.” In the first generation, the actor indeed lowered their head, but when they raised it, the AI produced a “hallucination,” turning the actor’s face into a cat. In the second generation, the action was correct, but the actor’s clothing color and style changed completely, breaking the continuity of the shot. In the third generation, the character, action, and clothing met the requirements, but the background for the frontal shot was an indoor room, while the reverse shot was set in a park, creating a complete mismatch.

“For that few seconds of footage, we adjusted and generated dozens of times before finally obtaining usable material.”

The core role in the AI video industry is the “card drawing master,” whose job is to repeatedly generate images and sift through dozens or even hundreds of results to find usable material, often leading to frustration.

Peng Yuhong noted that even with the same character and scene, merely changing the shot from close-up to wide shot can result in changes to the AI-generated background, character clothing, or even facial features. In back-and-forth dialogue shots, character expressions and lip-syncing often fail to connect, with issues like model clipping and prop misalignment being commonplace.

“Many shots require us to draw cards a dozen or twenty times to obtain usable material. Sometimes, even after dozens of attempts, we still can’t meet the requirements and have to simplify the script, abandoning complex shot scheduling in favor of close-ups to avoid mistakes.”

While the public perceives AI video as having extremely low costs, the reality is that costs depend entirely on the team’s professionalism and proficiency.

Industry insiders indicate that experienced teams can generate one minute of usable video for around 200 yuan, while teams still in the exploratory phase may find costs exceeding 500 yuan per minute, or even higher.

Peng Yuhong’s team, still in the exploratory phase, estimates that producing 30 seconds of video material costs about 800 yuan in computational expenses—this 30 seconds does not guarantee all usable footage. Liu Shuai’s team, while producing “Dilemma,” ended up with over 100 usable shots from more than 3,000 generated images, painstakingly selected and refined.

Despite this, the cost of AI video remains highly attractive. Currently, Seedance’s premium membership is priced at 499 yuan per month, including 15,000 points. Under discount conditions, generating a 15-second video costs only between 45 and 75 points, translating to a basic cost of just over 1 yuan for a 15-second video, which is unimaginable in traditional film production.

This extreme low cost has led to a mixed quality of content in the AI short drama and comic market. As the industry is still in a technical exploration phase, market differentiation is already evident. On one side are creators like Liu Shuai and Peng Yuhong, who aim to use AI tools to craft high-quality content with cinematic quality; on the other side are numerous teams targeting the low threshold provided by AI and engaging in a crude approach of “bulk content filling.”

Liu Shuai revealed that “currently, the outsourcing production quotes for AI short dramas have dropped to as low as 400 yuan per minute, with some teams using automation tools to achieve 800 to 1,000 minutes of comic content production in a single day.” Most of this content only seeks basic visual and dialogue coherence, entirely disregarding shot language, narrative rhythm, and content quality refinement, yet still gets accepted by major short drama platforms.

Behind this chaotic industry phenomenon lies capital logic and platform considerations.

An insider close to the Volcano Engine noted that “TikTok previously engaged with some short drama companies, hoping to use automated tools to produce AI short dramas and comics in bulk. ‘First, flood the market with low-quality content, then launch models like Seedance 2.0 and premium projects to create a contrast,’ to generate buzz.”

“Later, when Sora was shut down, this project was also halted.”

AI Has Become Cheaper, But Human Costs Have Increased

Regardless, AI has challenged traditional film production costs. “A few people and a few thousand yuan can create a blockbuster,” is no longer a dream.

In Liu Shuai’s view, AI has indeed lowered costs, which can be divided into two dimensions: explicit monetary costs and implicit risk costs. The reduction in explicit costs is evident: AI eliminates the need for real scene setups, equipment rentals, and large crew personnel expenses, compressing the heavy asset investments of real shooting into quantifiable computational costs.

However, the deeper cost reduction lies in controlling implicit risk costs. “Traditional film shooting is at the mercy of the weather. Once the crew arrives, changes in weather, actor health issues, scheduling conflicts, on-site safety accidents, or even post-production actor scandals can halt an entire project. Any uncontrollable factor can lead to project suspension.” AIGC digitizes the entire production process, transforming uncontrollable physical shooting into controllable digital generation, which is the core cost optimization AI brings to the industry.

On the flip side, AI has also introduced new cost calculations. Industry professionals indicate that while AI saves on filming costs, the core expenses of a production have never been on the machines but on the people.

Peng Yuhong calculated that producing a conventional AI live-action short drama with 60 episodes costs at least 300,000 yuan just for the production phase, excluding script adaptation and IP fees.

“Claiming that AI makes short dramas cheap only applies to content that does not pursue quality. If you want to create high-quality content that can compete with classics in the market, costs cannot be reduced significantly,” Peng stated. The time spent refining the first episode alone took over a week, involving repeated trial and error, adjustments, and invisible time and labor costs.

These investments often cannot be covered by revenue from distribution. Industry insiders report that “currently, the highest price for short dramas is about 2,000 yuan per minute, with a typical short drama lasting around 120 minutes, which totals only 240,000 yuan at best.”

Liu Shuai emphasized that AI lowers the production threshold, not the creative threshold. “Many believe that making videos with AI is as simple as inputting a few words and hitting confirm, but that is far from the truth. Professional script structure, shot language design, light and shadow aesthetics, and emotional communication still require skilled professionals to complete.”

“AI is like a high-performance sports car; ordinary people can only drive slowly on city roads. Only professional racers can fully unleash its potential.”

For this reason, the current labor costs in the AI video industry have not decreased but have actually risen. Traditional crew roles like production assistants, camera assistants, and lighting assistants are being replaced by AI; however, the demand for hybrid talents who understand content, aesthetics, shot language, and can proficiently use AI tools is increasing.

“The most expensive elements are not tokens or membership fees, but human emotions and thoughts,” Liu Shuai repeatedly emphasized. Technology can be quantified, but the creator’s delicate capture of emotions and the refinement of stories are truly priceless.

Peng Yuhong shares a similar sentiment. As a seasoned screenwriter, she initially thought her greatest advantage in entering AI short dramas was her scriptwriting ability, but she soon realized she needed to also take on the role of a product manager, overseeing the entire project process while repeatedly adjusting prompts with the team to address various unexpected issues in AI generation. “Team members often self-doubt, questioning whether their abilities are lacking or if the technology itself is immature, which can be a very frustrating process.”

In her view, while AI seems to allow a few people to complete a drama, it actually raises the demands on people. Practitioners must not only write scripts but also understand storyboarding, shot language, AI tool logic, and even computational scheduling, all of which impose far greater requirements on creators than traditional models.

However, it is undeniable that AI has opened a new door, enabling small teams and even individuals to realize their film dreams, and has brought content creation back to its creative essence.