Why an AI Animated Series Can Reach Eleven Seasons: From Generative Models to Production Systems

On August 15, 2026, the eleventh season of the AI animated series The Legend of Ten Thousand Demons Picture Scroll officially launched [1]. Previously, its tenth season had topped the Hongguo Trending Chart, with public data showing its heat index reached 80.6 million and its collection count reached 1.785 million [2]. Beyond its impressive chart performance, what is even more striking is that this series has continuously updated to its eleventh season.

Today, it is no longer uncommon for creators to use AI to generate a beautiful, few-second shot. The real challenge lies in achieving continuous updates around the same set of characters and worldview: character appearances must remain consistent, subsequent content must not conflict with existing plotlines, scenes, costumes, and sounds need efficient reuse, and failed shots require repeated generation and review. Continuous updates to the eleventh season at least indicate that the team has consistently and effectively solved these core problems, rather than starting from scratch each season.

Meanwhile, the production threshold for AI videos is rapidly decreasing. China Newsweek, citing relevant data from the China Netcasting Services Association, reported that in the first quarter of 2026, the entire industry launched approximately 128,000 micro-dramas, with AI micro-dramas accounting for about 122,000 of them [3]. However, this cost reduction has not automatically translated into profit. The same report noted that one team invested approximately 200,000 yuan to produce 11 AI animated series, with total revenue after one month online being only a few hundred yuan [3].

Considering these figures, a more thought-provoking question arises: When more and more people can generate videos at low cost, why do some projects quickly sink, while others can continuously update around the same IP?

Model capabilities alone cannot fully explain this difference; continuous production requires a stable and operational production system. This article does not aim to reconstruct the internal technical architecture of The Legend of Ten Thousand Demons Picture Scroll but uses this case as a starting point to explore why, as AI video moves from single generation to continuous production, teams need to simultaneously focus on whether models can generate the required shots and whether the system can reliably deliver.

How AI Video Goes from Story to Final Product #

To understand this transformation, we can start with the simplest generation method:

Write a prompt → Video model → A video.

This method is suitable for demonstrating model capabilities but struggles to support the production of dozens of continuous episodes. Actual production more closely resembles the following pipeline:

Story / IP → Script → Storyboard → Character and Scene Assets → Image and Video Generation → Voiceover and Music → Editing → Distribution → User Data → Next Production Round

This production chain indicates that the creator’s role has not disappeared, but their work content and focus have shifted.

Many AI animated series’ stories originate from web novels, existing IPs, or original screenplays. While Large Language Models (LLMs) can rapidly generate large amounts of text, what is truly scarce in the content industry is not the ability to “generate another 50,000 words,” but rather stories that can continuously attract audiences. In this process, LLMs are more suited to assist screenwriters in organizing character relationships, breaking down chapters, supplementing dialogue, and checking content continuity. If a story has already achieved considerable readership, collections, comments, or paid data, it at least indicates that the subject matter has successfully reached and retained audiences.

Text content then needs to be translated into audio-visual information. In a novel, it might suffice to describe:

Jiang Yuechu walked through Chang’an West Market. Street vendors’ cries rose and fell, then she suddenly paused—a string of copper bells under an eave jingled without wind.

This text is sufficient in a novel; readers would mentally fill in the market scene, character movements, and unusual atmosphere. However, during video production, the team must clearly define what the audience sees first, when they perceive something unusual, and through whose perspective they receive information. “She sensed something unusual” must also be translated into concrete clues that the audience can directly perceive (see and hear).

A simple paragraph might be broken down into the following shots:

The key here is not to divide a single sentence equally into five segments, but to carefully arrange the sequence of information release: first establishing space and character, then providing an unusual clue, followed by presenting the character’s reaction, and finally confirming the threat. Shot size, gaze, sound, and editing rhythm collectively complete this narrative segment.

This is Shot Breakdown. If the team needs to visually represent composition, camera angles, and action relationships through images, they will further create a Storyboard. Whether using a text table or a storyboard, the production team can describe shots using structured fields such as Shot ID, Narrative Function, Character, Scene, Action, Camera Angle, Duration, Prompt, and Reference Assets. Thus, a story is transformed into a set of assignable, executable, reviewable, re-triable, and acceptance-tested production tasks.

After the story is converted into structured shot tasks, the team also needs to convert them into inputs receivable by the model. In generative models, text prompts, character reference images, scene materials, action videos, and spatial layouts can all serve as conditioning inputs to constrain the generation results. Among these, influencing generation results by utilizing character images, scene images, reference videos, or sound materials can be called reference-based generation. Some video models already support multiple reference inputs [4][5][6]. Therefore, an asset library must not only store the files themselves but also record which characters, scenes, and shots the materials are suitable for, and how the system should pass these materials to the model.

Creators typically interact with creative platforms, while the underlying base models are responsible for content generation. If an enterprise needs to integrate generative capabilities into its internal system, it might call one or more models via an Application Programming Interface (API). The table below does not aim to compare models but illustrates the different roles played by platforms, models, and enterprise APIs.

Some new models have begun to simultaneously produce visuals, dialogue, and ambient sounds in the same generation [4][6]. While this capability is expected to shorten the voiceover and audio-visual synchronization process, teams still need to address issues like sound consistency, dialogue modification, and failed reworks within the same generation workflow.

Web platforms can already meet the needs of many ordinary creators. However, if a content company needs to produce hundreds or even thousands of shots daily, employees would still have to copy prompts one by one, download, and organize files, turning such repetitive operations into an efficiency bottleneck.

As production scales up, teams may gradually organize these operations into the following production pipeline:

Script Database → Automated Generation of Shot Tasks → Combination of Prompts, Reference Materials, and Other Conditional Inputs → Call Models to Generate Candidate Materials → Automatic Quality Check → Manual Review → Failed Tasks Regenerated → Enter Editing → Save Final Assets

The above process does not mean that AI video companies have universally deployed similar systems; it merely outlines a possible engineering path: as scale increases, video models will gradually evolve from independent tools into critical components within a production system.

Why Generating a Single Shot is Easy, But Continuous Production for Over Ten Seasons is Hard #

Generating a single beautiful shot is no longer a particularly difficult task today. However, when a team needs to continuously produce dozens of episodes or even over ten seasons of content, it faces a completely different set of challenges. The most intuitive question is: Can the same character always remain consistent?

If a team inputs “a young woman in white ancient costume” for each shot independently, the model is likely to generate varying facial features, costume details, and even age. To reduce this “drift,” production teams typically rely on a Character Bible, reference images, and character asset libraries.

From a computer system perspective, teams can understand this drift as a state management problem: how does the system maintain character and plot states across multiple model calls? If each shot is generated independently, the model will not automatically remember everything that happened in the previous shot. In computer system terminology, this shot-by-shot independent generation is akin to stateless calling.

This is similar to the concept of “world state” in games. When a character’s equipment, location, or plot stage changes, the game system must remember these state changes, rather than starting anew in the next frame. The same applies to continuous AI video generation: generative models produce new visuals, while the production system is responsible for saving the accumulated character, scene, and plot changes and passing their state to subsequent shots.

Continuous dramas require maintaining numerous states, for example:

Character State
├── Identity
├── Appearance
├── Costume
├── Voice
├── Personality
├── Current Location
├── Character Relationships
└── Plot History

If a character changed clothes in the previous scene, they cannot suddenly revert to their old attire in the next shot; if a sword was broken in the previous episode, subsequent shots cannot restore it to its original state due to regeneration; if two characters have learned a secret, they cannot rediscover the same secret in the next episode.

Therefore, the difficulty in continuously producing AI videos lies not only in enabling a single model to generate more exquisite visuals, but also in the system’s need to maintain the narrative continuity of the story world.

Model demonstrations can also easily obscure another critical issue: the availability rate of generated results. Commercial production is not solely concerned with what a model can achieve at its best; teams are more focused on how many times, on average, generation is required to obtain a usable shot for the final cut.

Suppose a micro-drama requires 100 shots, and each shot, on average, needs to be generated 5 times. The team would then need to call the model approximately 500 times, ultimately using only 100 results. Roughly calculated, the availability rate of candidate results is only 20%. At this point, the price of a single API call can no longer fully reflect the true cost. Teams need to focus more on the usable shot rate and the actual cost required to obtain a single qualified shot. This article refers to the latter metric as the cost per qualified shot.

The cost per qualified shot includes not only model inference fees but also expenses for failed retries, manual review, re-editing, waiting time, and subsequent rework. Even if a model vendor halves the price per generation, as long as the average number of retries does not decrease proportionally, the final production cost may not decrease accordingly. For production teams, the cost per qualified shot is often more practically significant than the single generation price announced by model vendors.

Upon entering commercial production, teams also cannot merely compare “which model generates faster.” Teams must simultaneously consider Latency, Throughput, and cost per qualified shot. If a director is revising a critical shot, generation latency is more crucial; whereas if a team needs to batch-generate 500 ordinary shots overnight, throughput and cost per qualified shot might be more critical.

Therefore, teams should not limit themselves to asking “Which is stronger, Kling or Veo?” but rather ask: Under the current shot, quality requirements, time, and budget constraints, which model is most suitable? Model selection must be subservient to the overall objectives of the entire production system.

As Workflows Become More Complex, Who Decides the Next Step #

Engineering practices for machine learning systems have long revealed similar issues. Google researchers pointed out in 2015 that model code often constitutes only a small portion of a production system; beyond the model, there are numerous components including data, configuration, resource management, services, and monitoring [10]. Subsequently, Machine Learning Operations (MLOps) practices further integrated version management, pipelines, deployment, metadata, and monitoring into a unified management scope [11].

While AI video production cannot directly adopt MLOps, teams will encounter similar problems as production scales. For example, if a shot needs to be regenerated six months later, the team cannot just save the final video; it also needs to retain a complete generation record:

Shot #0231

Script Version:       v12
Character Asset:      JiangYuechu-v4
Scene Asset:          ChanganStreet-v7
Prompt Version:       v18
Video Model:          Model-A / Version-X
Reference Images:     ref-03, ref-08
Retry Count:          3
Output:               candidate-17
Review Result:        Accepted

These records enable the team to reproduce the generation process of the shot. Otherwise, the team would soon wonder: “This shot looks excellent, how exactly was it generated?” If the model version, prompts, reference assets, and parameters are not recorded, the answer might simply be “I don’t know.”

Therefore, scalable production demands version control, asset tracking, task orchestration, and operational monitoring. However, fixed workflows cannot solve all problems. Teams still need to make judgments that cannot be predefined by rules: for example, if a character’s appearance “changes,” should the prompt be modified, the reference image replaced, or another model used? If the budget is more than half consumed, is this shot still worth trying?

Such judgments are better handled with the assistance of Agentic Systems. I tend to view agents as a scheduling and control layer on the production line, rather than a “virtual director” capable of fully automating all work from script to final cut.

If a team only generates ten candidate shots per day, humans can check them one by one; but if thousands or even tens of thousands of candidate results are generated daily, quality checking will become a new capacity bottleneck. The system can first use automatic quality inspection to screen for issues like character identity drift, obvious visual anomalies, subtitle errors, audio-visual desynchronization, and storyboard deviation, then pass the pre-screened results to human reviewers. Missed detections will increase human workload, and if the system misidentifies usable material as unqualified, regeneration attempts will also increase. Both types of errors affect the cost per qualified shot. Automatic quality inspection cannot replace a director’s judgment of whether a work is natural or aesthetically pleasing, but it can significantly reduce the time spent by humans checking obvious failed results.

When designing such systems, the key is not how many agents are deployed, but rather clarifying three types of responsibility divisions: deterministic rules and workflows for tasks with clear boundaries, agents for situations with clear constraints but which cannot be exhaustively predefined, and important decisions affecting quality, cost, and risk to be made by humans.

Tasks like asset querying, budget checks, and “escalate to human after three failures” are suitable for deterministic code execution. Agents are better suited for tasks where inputs are not completely structured, require contextual understanding, but are still constrainable—for example, determining if a shot failure is due to character consistency issues, and whether the next step should be to change the reference image or the model. As for important decisions like whether to release a work, exceed the budget, handle copyright risks, and accept the final cut, these still require human approval.

For agentic systems to truly enter a production environment, they cannot solely rely on lengthy conversations; each shot needs clear structured fields, such as:

shot_id
character_ids
scene_id
reference_assets
model
retry_count
review_status

This allows the system to execute explicit rules: for example, hand over to human when retry_count >= 3; prohibit entry into the release process when review_status != approved.

The system also needs to retain complete Tracing records, including: which agent made the judgment, which model was called, at what step it failed, how many times it was retried, and when it was escalated to a human. These records help the team troubleshoot errors and can be used to calculate production metrics such as usable shot rate, average retries, human intervention ratio, and actual costs. Only by recording structured states and the complete execution process can an agentic system move beyond the realm of “several agents conversing and coordinating work” and truly become an observable, measurable, and optimizable production control layer.

Some agent frameworks already provide these capabilities. For example, the OpenAI Agents SDK supports tool calling, structured output, task handover, constraint checking, human approval, and tracing [12]. This is merely an implementation example; what’s more important is the underlying general design pattern, not a specific SDK.

How Data Influences the Next Production Round After Content Launch #

After production is complete, teams also need to publish their content on Hongguo, Douyin, or other platforms. Platforms do not evenly display micro-dramas to all users; recommendation systems determine who sees the content first, how much exposure it receives, and whether the platform expands its distribution.

For example, when a certain type of “reincarnation” micro-drama is launched, the platform will first give it some exposure; if the completion rate or interaction data performs well, the platform may continue to recommend it. Production teams then create more similar works based on data performance, and the platform consequently gains more similar content candidates. Superficially, users appear to increasingly prefer this genre, but this conclusion requires careful consideration.

Clicks, views, and completion rates are all Implicit Feedback. These behaviors cannot be directly equated with users’ inherent preferences, because users can only interact with content they have actually been exposed to. The recommendation system determines which content receives exposure, and this act itself changes the distribution of subsequent data, which is the problem described by Exposure Bias.

Research has shown that the feedback loop formed by “recommendation-interaction-retraining” may further reinforce Popularity Bias, causing content that has already received more exposure to continue to receive more, reducing the diversity of recommendation results [13][14].

In AI video production, this mechanism will form a new closed loop:

Recommendation system identifies a content type performing well → Platform increases exposure → Production team creates more similar content based on data → Recommendation system gets more similar candidates → Users see even more similar content

While this closed loop can improve short-term production efficiency, it may also drive the convergence of themes, characters, plots, and visual styles. In the long run, the “user preferences” observed by platforms will be more constrained by the platform’s past exposure distribution strategies, and not equivalent to users’ free choices among all content.

Therefore, when content teams claim “data tells us what users like,” a more accurate conclusion should be: Under current recommendation and exposure mechanisms, this content has achieved better behavioral feedback.

Previous recommendation systems primarily addressed the problem of “what to show users from existing content.” The integration of generative AI means platform data can further influence “what the next batch of content should be produced.” Recommendation systems may therefore no longer solely reside downstream of content distribution but indirectly influence upstream production decisions through data feedback.

In the future, agentic systems may also effectively connect data analysis, theme suggestions, small-scale pilot production, traffic testing, and feedback loops. This is currently still an engineering extrapolation, not a widely deployed industry reality. What truly needs to be discussed in advance is: which metrics are suitable for automatic optimization, which feedback is already influenced by exposure mechanisms, and which content decisions cannot solely be determined by short-term data.

What Becomes Scarcer as Generation Costs Decrease #

Some AI video teams report that production costs are rapidly declining. In an interview with China Newsweek, one team stated that the production cost of a 60-episode AI animated series decreased from approximately 80,000 yuan to 20,000 yuan within several months; another team mentioned that the production cost for some AI-simulated live-action dramas could be compressed from hundreds of thousands or even nearly a million yuan to about 80,000 to 100,000 yuan [3].

These figures come from specific company projects, and there is no unified standard for defining “production cost,” so they cannot be directly generalized to the entire industry. However, these cases at least show that AI technology can significantly reduce the trial-and-error costs for certain projects.

In some projects, production teams might have previously invested hundreds of thousands or even millions of yuan to verify if a story had a market and an audience. Today, the same budget can test more projects, and more competitors can also use this method for trial and error.

When production costs decrease, and more entrants join the market, while the time users can spend consuming content does not increase proportionally, the attention each work can receive will inevitably be squeezed. China Newsweek reported that one team produced approximately 30 works per month, with only one eventually becoming a hit, and that hit precisely covered the losses of the other 29 works [3].

This case at least shows that generative AI has not made the distribution of content revenue more even; a few blockbusters can still cover the losses of many ordinary works. AI lowers trial-and-error costs but also attracts more teams to the field. The closer generative capabilities get to infrastructure, the more likely competitive advantages will shift beyond the models themselves.

Stories and IP remain crucial. When users binge-watch a series, they are usually concerned with the characters’ evolving fates, not whether the shots were generated by Kling, Seedance, or Veo. Teams continuously accumulate character, scene, and sound assets, and meticulously record character relationships and story states. Once this content is solidified into reusable production assets, it not only reduces the design costs for subsequent series but also can build unique competitive advantages.

A stable production system will also become a key competitive advantage. When all competitors can call similar models, the differentiation will be more reflected in the usable shot rate, number of reworks, speed of failure localization, model selection, and manual review costs. This is quite similar to industrial machine learning systems: the best model does not automatically equate to the best system.

Furthermore, there is a critical capability often obscured by the “data-driven” philosophy: judging what data truly means. Is a high completion rate due to the story itself being more engaging, or because the platform pushed it to more matching users? Is a low click-through rate due to lack of market demand, or issues with the cover and exposure placement? If a certain theme suddenly becomes popular, has a long-term preference shifted, or has the recommendation system amplified the influence of a certain type of content in the short term?

Data and agents can both aid content production, but mature teams must still judge what conclusions data supports and when correlation should not be directly interpreted as user demand.

As generative capabilities become increasingly commoditized, the value of “who is best at writing prompts” may no longer be paramount. Good stories, reusable assets, stable operational production systems, and a true understanding of the audience may hold more long-term value.

Future Outlook #

What is truly worth watching in the next phase of AI video might not be how much the next generation of models has improved in image quality, but rather how video generation is gradually evolving from an independent capability into an organic component of content production infrastructure.

I discussed this change in my previous article How Generative AI is Reshaping Recommendation Systems: in the past, people were more concerned about recommendation systems influencing “what we can see”; the new change brought by generative AI is that it begins to influence “what the content we see actually looks like.”

From the production side, this change presents another aspect: when generative systems and recommendation systems become increasingly connected, platforms not only determine which content gets exposure, but user behavior can also, in turn, influence the production of the next batch of content. Teams may tightly connect theme selection, pilot production, generation, distribution, feedback, and reproduction stages, forming a faster feedback loop.

While this closed loop can improve production efficiency, it may also lead teams to increasingly rely on short-term completion rates, click-through rates, and interaction data, leading to the repeated replication of successful themes and formats. Systems may become increasingly adept at predicting which content is likely to receive feedback, but teams still need to carefully judge: what content is truly worth producing, what projects are worth continuous investment, and what works are worth audiences investing their precious time to watch.

References #

[1] Hongguo Short Drama Official Weibo. The Legend of Ten Thousand Demons Picture Scroll Season 11 Launches on August 15. 2026-08-11.

[2] Annette. Hongguo Trending Chart TOP5: AI Short Dramas Dominate, Liu Xiaoxu Applies to Join the Battle!. Sina News, 2026-08-12.

[3] Wang Shihan. Micro-Drama Shakeup: 95% Are AI, But Earning Money is Like Winning the Lottery. China Newsweek, 2026-06-16.

[4] ByteDance Seed. One-take Creation, Flexible Referencing: Introducing Seedance 2.5. 2026-07-31.

[5] Kuaishou Technology. Kuaishou Kling AI Unveils “Multi-Image Reference” Feature to Further Tackle AI Video Consistency Challenges. 2025-01-23.

[6] Google DeepMind. Veo. Accessed on 2026-08-22.

[7] MiniMax. MiniMax Hailuo 2.3: A New Level of Complex Video Performance. 2025-10-28.

[8] Alibaba Cloud. Model Studio: Video Generation. 2026.

[9] Runway. Introducing Runway Gen-4.5. 2025-12-01.

[10] Sculley, D., Holt, G., Golovin, D., et al. Hidden Technical Debt in Machine Learning Systems. Advances in Neural Information Processing Systems 28, 2015.

[11] Google Cloud. MLOps: Continuous Delivery and Automation Pipelines in Machine Learning. Google Cloud Architecture Center.

[12] OpenAI. OpenAI Agents SDK Documentation: Agents, Agent Orchestration, Guardrails, Human-in-the-loop and Tracing. Accessed on 2026-08-22.

[13] Mansoury, M., Abdollahpouri, H., Pechenizkiy, M., Mobasher, B., & Burke, R. Feedback Loop and Bias Amplification in Recommender Systems. CIKM 2020.

[14] Gupta, S., Wang, H., Lipton, Z., & Wang, Y. Correcting Exposure Bias for Link Recommendation. Proceedings of the 38th International Conference on Machine Learning, 2021.