Kling 4.0 Hands-On Test: AI Characters Enter Live-Action Exhibition-Walkthrough Videos — How Far From One-Stop Finished Video?

By | Singularity Research Society, author | Xu Ting, editor | Dr. Qi

Late at night on October 30, Kuaishou published an announcement on the Hong Kong Stock Exchange: its video generation business Kling completed the capital increase delivery, with all 19.802 billion yuan in capital increase funds received, equivalent to about 2.905 billion US dollars. Kling's post-investment valuation is about 17.9 billion US dollars, equivalent to more than 120 billion yuan, higher than Kuaishou's Hong Kong stock closing market value of about 96.8 billion yuan that day.

"One Kling is worth more than all of Kuaishou," especially considering that in the first half of 2026, Kling's revenue accounted for only 2.17% of Kuaishou's total revenue. What the capital market is buying is clearly not Kling's current revenue, but its future in the AI video track.

No matter how grand the capital story is, it ultimately still has to return to the creative scene to be tested. Recently, we used Kling 4.0 to produce an exhibition exploration video. All live-action on-camera segments were handed over to AI, combining actual exhibition footage with AI characters, splitting the generation into 7 groups of shots, and then completing subsequent production through features such as synchronized audio and video, video modification, and timeline. Before testing, we prepared about 3,000 words of prompts to clarify subject relationships, camera movement, composition, and sound requirements as much as possible, reducing instruction ambiguity.

The actual test results were mixed. In some shots, character integration, camera movement, and composition were already quite natural, with almost no obvious incongruity; but in other clips, the AI would replace the robot on stage with the host off stage, turn a one-take shot into several hard cuts, and even after modifying the video, turn originally normal sound into a string of unrecognizable garbled noise.

AI can already generate many stunning shots, so why does the final cut still require repeated trial and error? How far is Kling 4.0 from "one-stop finished video," and what key capabilities is it still missing? Under a 18 billion US dollar valuation and industry resource restructuring, can stable delivery become the commercial watershed for AI video?

Kling 4.0 upgrades again, from generating images to executing creative ideas

In the large model competition among domestic internet companies, Kuaishou chose a relatively focused route: it did not put its main chips on general-purpose chatbots, but continued to invest in video generation. This is closely related to Kuaishou's original business ecosystem. Advertising, e-commerce, short dramas, film and television, and creator content all have massive video production needs, and model capabilities also have the opportunity to be converted into revenue through membership subscriptions, enterprise services, and API calls.

This is a business closer to money, and also more money-burning. The paid scenarios for video generation are relatively direct, but training, inference, computing power, and talent are all expensive. Kling thus gradually moved from an internal technical project at Kuaishou to the company's strategic center.

This shift in weight is clearly written in organizational and capital actions. In April 2025, Kuaishou established the Kling AI business unit, with its head reporting directly to Cheng Yixiao. In July, Kling advanced an independent spin-off financing arrangement, with a financing scale of nearly 3 billion US dollars and a post-investment valuation of about 18 billion US dollars. Tencent, Alibaba Cloud, Baidu, as well as multiple local industrial investment funds and entertainment industry parties entered the shareholder list. On September 30, Yu Yue, former head of the community science line, was transferred to serve as director and CEO of Kling AI Company, responsible for operations and management. On October 30, this round of independent financing officially completed delivery, with actual capital increase funds received of about 19.802 billion yuan, equivalent to about 2.905 billion US dollars.

From a first-level business unit to independent financing, and then to a dedicated CEO, Kling is gaining the resource allocation of an independent company and also beginning to face the operating test questions of an independent company.

Commercialization data is its most weighty bargaining chip at present. According to data disclosed by Kuaishou, in the first quarter of 2026, Kling AI revenue exceeded 650 million yuan, a year-on-year increase of more than 300%; second-quarter revenue exceeded 850 million yuan, a year-on-year increase of more than 200%, with the first-half total exceeding 1.5 billion yuan.

But revenue growth does not mean the business model is already worry-free. As business independence strengthens, it also needs to answer several more realistic questions, such as whether revenue growth can continue, how computing power costs will be absorbed, whether enterprise customers are willing to continue renewing, and whether the product can build sufficiently solid competitive barriers, among others.

Reports that Kling is preparing for a Hong Kong IPO have appeared in the press, but related arrangements are still at an early stage, and there is room for adjustment in listing timing and scale. Capital has given Kling greater room, and also raised its passing line. Next, the market is asking not only what the model can still generate, but also whether technical highlights can turn into sustained returns.

Kling 4.0 arrives precisely against this backdrop, and is also a product node where Kling attempts to take another step toward professional creative workflows.

According to official Kuaishou information, the Kling 4.0 upgrade focuses on realism, controllability, and professionalism: a single native generation can last up to 30 seconds; reference materials such as images, videos, subjects, and timbres can be flexibly combined, supporting up to 15 reference materials and 10 keyframes; and the prompt length limit has also been further raised. Officially, these upgrades are intended to better handle high-dynamic scenes, complex camera movement, and continuous narrative.

These capability changes correspond to a shift in the evaluation standards of the AI video industry. Early models first had to prove whether they could convert textual descriptions into a segment of footage: whether character actions were natural, whether the image was realistic, and whether lighting and shadow were close to real photography. As generation quality improved, creators began to make more complex demands: characters must remain consistent across shots, the camera must move according to specified shot size, composition, and rhythm, and multiple clips must also maintain basic spatial relationships and narrative continuity.

Beyond “what is in the frame,” the model also needs to handle what role each piece of material plays, which elements can be changed, which information must be preserved, and how the creator wants the shot to unfold. Kling 4.0's upgrades to reference materials, camera movement, and generation duration are precisely a response to these needs.

Product form is evolving in tandem with model capabilities. Kling is gradually integrating generation, editing, canvas, and timeline capabilities into a single workspace, hoping to reduce the cost of users switching between different tools. The Seedance series previously launched by ByteDance's Seed team also treats long-form narrative, multimodal references, and video editing as important directions for product iteration.

Each company's approach differs, but the competitive goal is becoming clear: what creators expect is no longer just a button that can generate clips, but a set of tools that can help complete a work.

If the first half of AI video was about generation capability, the second half tests the ability to execute creative intent. But do longer generation durations, more reference materials, and richer editing capabilities mean creators can more accurately control the final result?

Qidian conducted a hands-on test using an exhibition-coverage video. The exhibition venue was live-action footage, and AI characters needed to enter the existing environment and complete different camera movements; across multiple clips, the relationships between characters and scenes also had to be maintained and connected with voiceover, sound, and post-processing.

Kling 4.0 hands-on test: between a good shot and a good finished video, how many “redos” are there?

This test used Kling 4.0 throughout. The video combined live-action footage with AI characters, was split into 7 groups of shots, and generation, modification, and material organization were completed in the canvas. To minimize instruction ambiguity as much as possible, we used an AI assistant and a colleague from the team's editing staff to organize about 3,000 characters of prompts, specifying subject relationships, composition, shot size, camera movement, and sound requirements.

In other words, this was not a random test that completely left the model to its own devices, but a real-world exercise with professional assistance and a clear delivery goal.

Kling 4.0's upper limit is indeed surprising enough. The best-performing clips could accurately complete the camera push-ins required by the prompts, and composition, shot size, and cut rhythm basically met expectations. Even more rare, the integration of AI characters with the live-action environment was quite natural; the characters did not have an obvious “pasted onto” the background feel, and they did not seem abrupt when interacting with the environment.

Such footage is already close to a state that can be delivered directly. It shows that Kling 4.0 can not only generate a few sample clips suitable for demonstration; under the right conditions, it can indeed produce content with professional use value.

Sound was also a relatively bright spot in this test. Most ambient sound and voiceover could be used directly, and the overall usability rate was higher than that of the visuals. Simultaneous audio-video output reduced the step of dubbing each segment one by one, which has clear efficiency value for creators who need to quickly complete a short video.

But the highlight performances could not be reproduced consistently. The first reason is that the model still deviates in handling the relationship between subjects and scenes. In one generation, the host who was supposed to stand below the stage directly replaced the robot on the stage;

In another clip, the AI replaced the host in the background video with a character. Even after adding relational descriptions such as “watching,” the result was not noticeably corrected. In other clips, the proportions of characters and scenes were unbalanced, suddenly giving a normal exhibition visit the look of “accidentally wandering into Lilliput.”

The core of these bugs is not image quality, but whether the model truly understands “who is who, who is where, and who interacts with whom” in the frame.

Second is long takes and continuous narrative. We specifically designed a set of one-shot instructions, but after several attempts, what we got was still hard cuts between different scenes. Although the character kept walking, the camera did not form continuous movement, and the surrounding environment was also substantially modified. The restrictions in the prompt on the use of materials were not stably executed. Longer duration does not naturally equal an upgrade in narrative capability; the model can generate longer content, but it may not be able to continuously maintain spatial, character, and camera relationships.

Finally, there is consistency in modifications and sound. After some clips were modified through video editing, not only did they fail to be corrected according to instructions, but subtitles were added without authorization; originally normal sound could also turn into garbled text in subsequent generation, voices preconfigured for the character subject were not stably referenced, and the same character in different clips could have different vocal timbres. The canvas and timeline did improve material management, but BGM, transitions, sound effects, and the modification workflow still have not formed a sufficiently smooth closed loop for finished videos. If anything goes wrong, everything has to be “pushed back and redone.”

This is also the most direct impression from this hands-on test: Kling 4.0 has raised the upper limit for generating good shots, but stably delivering a video that meets requirements still requires creators to backstop the key steps.

The costs incurred during this process are easily obscured by the price of a single generation. Prompt debugging, repeated generation, material screening, failed modifications, waiting time, and manual inspection and repair together constitute the real production cost of the video. For individual creators, they consume time and energy; for advertising, short-drama, and film and television teams, they also affect schedules, budgets, and delivery acceptance.

Judging from the test results, Kling 4.0 can already generate footage with professional use value in some scenarios, but there is still room for improvement in stably executing complex requirements, preserving effective content during modifications, and smoothly completing the entire production process. These issues also bear on the focus of competition in the next stage of the AI video industry.

Products such as Kling, Seedance, Vidu, Hailuo, and Tongyi Wanxiang are continuously iterating around image quality, character consistency, continuous narrative, generation cost, and vertical scenarios. A single version upgrade can bring about phased capability changes, and as foundational generation capabilities keep improving, it is increasingly difficult to fully measure a product's actual value just by whether a particular sample video looks impressive.

Kling's commercialization progress already has relatively clear public data, and its revenue growth has also made it a player worth continuing to observe in the video generation field. However, revenue scale, user growth, profitability, and long-term competitiveness are different dimensions, and without competitor data on the same basis, it is also difficult to assert on that basis that it is already comprehensively leading in commercialization.

Competition is likewise taking place in the application ecosystem. ByteDance has products and services at different links such as Seedance, Jimeng, Jianying, Douyin, and Volcano Engine; Kuaishou also has content platforms and Kuaijing. After obtaining more independent financing and operating space, Kling needs not only to continue coordinating with its parent company but also to build its own product ecosystem for external developers and enterprise customers.

Recently, reports said that Xiaoyunque, ByteDance's AI short-drama creation platform, was merged into Jimeng; the product still retains an independent entry point, and the related short-drama capabilities have been integrated into Jimeng. This forms an interesting contrast with Kling's independent financing: one side seeks independent operating space for a single business, while the other further integrates the capabilities and operations of similar products. These two organizational choices cannot directly prove which path is superior, nor are they enough to represent a unified trend across the entire industry, but both point to the same real problem: enterprises need to rejudge which capabilities should be developed separately, which products should be integrated, and which links can create sustained commercial value.

As more and more third-party creation platforms can access different video models, merely possessing generation capability may no longer necessarily become an application's exclusive advantage. Users can compare the results, speed, and price of different tools, and can also switch models according to the task. For an application to retain users, it needs to provide value beyond the model: a workflow better suited to a certain creative task, more refined editing and collaboration capabilities, more reliable content management, or a direct connection with distribution and commercial conversion.

The first is generation stability. For the same task, can it relatively stably generate results that meet the requirements? Do successful frames appear by chance, or can they be repeatedly reproduced under similar conditions? For professional users, a stunning result is certainly useful, but stability is what determines whether it can become a routine tool in the production process.

The second is the controllability of modifications. When there are deviations in characters, actions, or voices, can the tool complete corrections while preserving other effective content? If every modification may introduce new errors, the model may appear to provide more editing capabilities, but in practice it may increase rework.

Finally, there is the comprehensive delivery cost. In addition to the time and credits required for model generation, the investment in prompt preparation, material screening, failed retries, manual repair, and audio-visual post-production should also be calculated. Only by putting these costs together can one judge how many resources a usable video actually requires, and only then can one compare the efficiency of different tools in real workflows.

These three dimensions are further connected to commercial returns. Stable generation and modification capabilities can reduce repetitive labor; a smoother workflow can allow models to enter professional production scenarios such as advertising, short dramas, and film and television; and a predictable unit delivery cost affects whether customers are willing to use and renew for the long term.

For model companies, user scale, ARR, and financing valuation remain important, but enterprise renewals, API usage, computing power costs, and final profit also determine whether growth can be sustained.

In this process, mixing multiple models and tools is likely to persist for a long time. Creators can use one model to generate visuals, another tool to process sound, and then use professional software to complete editing and post-production. Model vendors must strive to get their capabilities into the production processes that users run repeatedly.

The $18 billion valuation prices Kling's technical potential and commercial prospects, and also provides space for its subsequent development. The 4.0 capability upgrade demonstrates its efforts in the direction of professional creation. What needs to be further verified next is whether these technical capabilities can be transformed into more stable results, a more controllable modification process, and delivery costs that customers can accept.

Good shots attract people; stable delivery retains them.