Hands-on test of Kling 4.0: water-sleeve drum dance, Singin' in the Rain, and a 30-second slice-of-life short

Hands-on test of Kling 4.0: Why is it so hard to make AI dance? Leiphone October 26, 2026 11:34:07 From Guangdong Updated October 26, 2026 11:55:09

English text is being prepared.

English text is being prepared.

In 2004, Zhang Yimou filmed an unforgettable water-sleeve drum dance in House of Flying Daggers.

Zhang Ziyi's dancer stands at the center of a drum array, leaping and spinning with the music as two long water sleeves flutter through the air, sometimes tracing huge arcs, sometimes whipping toward the drumheads around her. The performer's movements, the trajectories of the silk, and the dense drumbeats come together in what remains one of the film's most recognizable scenes.

This performance is hard to replicate to begin with. The dancer, the water sleeves, and the drum array must coordinate perfectly, and the film can also rely on camera staging and editing to make every drumbeat look crisp and clean. Now condensing all these requirements into a single prompt and having AI generate 15 seconds in one go sounds like asking a bit too much of it.

But we still wanted to try. After all, one of the key upgrades in Kling 4.0 is the generation of complex dynamics.

On October 15, Kling 4.0 officially launched, extending native generation length to 30 seconds and bringing updates in complex dynamics, reference control, and video editing. Besides this water-sleeve drum dance, we also brought in the classic solo from Singin' in the Rain, had it shoot a 30-second slice-of-life short of an ordinary young woman strolling through an old neighborhood, and tried to modify an already-generated video.

In this hands-on test, we wanted to see how far Kling 4.0 could take these requirements, and whether it is really as effortless in actual use as the official sample videos make it look.

01

The water-sleeve drum dance is inherently a highly difficult performance, and in the film it also has to balance the performer's body movements, the motion of the fabric, the positions of the drum array, and the camera staging. For a video generation model, it needs to handle how to make this whole chain of actions hold together in the same scene.

We originally thought the truly difficult part would be the dancer's midair flips, but after several rounds of generation, what turned out to be the biggest headache was how to make her actually hit the drumheads correctly.

We first prepared reference images for the character's look and the drum array, and instructed Kling about the red dance costume, the light pink long water sleeves, and the movements of spinning, shifting, leaping, and whipping the sleeves to strike the drums, hoping to reproduce the main features of this dance with a single 15-second generation. Autoplay

In terms of visual impression, this version performed reasonably well. The red-clad dancer's look basically holds up, the silk shows stretching and swirling dynamics in the air, and the whole scene is already evocative of the original film. But when it comes to the actual drumming, the dancer directly touches the drumheads with her hands, and the water sleeves sometimes merely flutter along with her arms without serving the function of striking the drums.

We then repeatedly adjusted the prompt around this issue, explicitly requiring that the drums must be fixed and that the tips of the sleeves are the only part allowed to touch the drumheads. Considering that the model might not understand the complete choreography, we also split the 15 seconds into several stages—the opening pose, the midair flips, and the continuous drumming—hoping the latter half would show clear changes in speed and force. Autoplay

After nearly seven or eight rounds of readjustment and testing, we still could not achieve the goal of striking the drums with the ends of the water sleeves throughout.

In some versions, the dancer still uses her hands instead of the water sleeves to strike the drums; in others, although she whips the long sleeves, she merely stands in place and turns slowly, lacking the expected leaps and spatial displacement. Even stranger, in some shots the drum bodies move on their own in order to achieve the goal of "striking the drums" stated in the prompt.

Of course, this is not a strict professional choreography test.

We did not draw shot-by-shot storyboards the way professional film practitioners do; the success of a film scene from 20 years ago itself has much to do with editing, storyboarding, and visual effects. This does not mean the model is incapable of completing this video—it only shows that, under a creative approach relying on reference images and natural-language descriptions, even if the same requirement is repeatedly emphasized, complex actions may not reliably be corrected in the intended direction.

02

By contrast, the Singin' in the Rain test showed us what Kling is better at.

Released in 1952, Singin' in the Rain left behind one of the most famous song-and-dance scenes in film history. Gene Kelly, in a suit and bowler hat, dances alone with an umbrella on a rainy-night street, spinning and tapping along with the music, and finally splashing into a puddle by the roadside. The coordination of the performer's movements, the musical rhythm, and the camera turns an ordinary street into a performance space full of charm.

This time we provided four reference images from the classic sequence and asked the model to generate a rain solo with the texture of a vintage Hollywood musical, retaining the black umbrella, the bowler hat, the streetlamp, and the wet street, while arranging movements such as turning, swinging the umbrella, tapping, and stepping in water.

The first version was quite good. The camera stays focused on the protagonist throughout, the brick wall and street behind him show no obvious scene mutations, and the rainy-night colors, lighting, and the texture of the soaked suit are all fairly natural. The jazz music and the final water-stepping action also basically match the rhythm, giving the whole clip a bit of the flavor of an old-fashioned musical film.

We especially liked a detail that is easy to overlook: as the character dances under the streetlamp, a shadow that changes with his movements appears on the wall behind him. This makes the image more than just a person dancing in the rain—it also establishes a connection between the character and the environment. Autoplay

However, Kling clearly has not yet taken care of every detail. At one point the male lead tilts his head back, and his hat falls off; when he raises his head again, the hat is back as if nothing had happened. In the final few stomps into the puddle, the splashes also look too regular, almost always spreading symmetrically to both sides from the same center point. The smile on the male lead's face is likewise a bit too perfect, lacking the natural shifts in expression that occur when a real person keeps moving.

Visual quality has always been an advantage Kling is relatively recognized for, but as Seedance keeps iterating, a beautiful image is no longer any one company's exclusive skill.

In July this year, Seedance 2.5 was the first to support 30-second generation, and upgraded multimodal references and video editing. Less than three months later, Kling 4.0 launched, with its update priorities likewise falling on 30-second generation, complex motion, and reference control. Some creators' same-prompt tests recently have also produced differing assessments: some prefer Seedance's realism in its characters, while others more approve of Kling's motion performance. Each has its strengths, and it is already hard to judge which is better based on a single pretty sample clip.

For Kuaishou, this competition also carries commercial pressure. In the second quarter of this year, Kling AI's revenue already exceeded 850 million yuan, up more than 200% year on year. What it needs to retain is not just users trying it out, but also paying creators who actually use it to produce content such as ads and short dramas.

And for these people, once the video is generated, the work often has only just begun.

03

Just want to change a suit—can the whole video be preserved?

The Singin' in the Rain generation turned out fairly well, so we wanted to try Kling 4.0's upgraded video editing feature. If a video's motion, composition, and scene are all already satisfactory, and we just want to change the clothes the character is wearing, can we avoid starting the gacha over from scratch?

We used the previously generated Singin' in the Rain as the original clip and made a simple request: change the male lead's dark suit to wine red, keep the original style, and do not change the character's face, dance movements, hat, umbrella, or the scene and sound.

Image before editing

Image after editing

In the video modified by Kling 4.0, just as in the original video, only the male lead's clothes in the original video were replaced. It is worth noting that although the video editing feature was used, the credit consumption for this modification was not much different from regenerating a video.

To add a point of reference, we also handed the same video to HeyGen Video Agent with exactly the same request.

The result was a bit unexpected: the suit did turn wine red, but the male lead was no longer the same person, and even the rainy-night street scene changed along with it. The dance segment we originally wanted to keep was also not fully preserved.

After further inquiry, HeyGen explained that the currently used Video Agent workflow can trim the original clip, add text, or regenerate materials, but it cannot yet modify only the suit in an uploaded video while keeping the rest of the image unchanged.

Here we need to distinguish two kinds of "editing": shortening a shot, swapping in a piece of footage, or adding subtitles belongs to arranging video content; changing the clothes of a person dancing in the frame requires modifying the original image and preserving other details through continuous motion. HeyGen Video Agent supports the former kind of operation, but in the entry point we used, it did not complete the latter task.

In our previous hands-on tests, we also went through many rounds of gacha, and in commercial production this experience is not just a matter of personal patience. When Yicai previously interviewed Mansen Culture, its founder Huang Haorong revealed that the team needs about 10 people to produce a premium AI short drama, about half of whom are responsible for gacha. According to the team's data at the time, it took an average of 5 generations to get one satisfactory image.

This time Kling successfully completed the suit color change, at least preserving the performance we were already satisfied with, which is worth acknowledging. But when the credit consumption for one edit is still not much different from regenerating, when the cost of modification will truly come down is probably an unavoidable question for these video models going forward.

04

30 seconds, an ordinary girl's life vlog

Beyond difficult actions, we also wanted to try another upgrade in Kling 4.0: native video generation of up to 30 seconds. Seedance 2.5 had already supported this duration before. Compared with Kling's previous generation's maximum 15-second generation capability, it still means creators can try to fit more content into a single generation.

English text is being prepared.

We wanted it to have a bit of the visual texture of an old DV; the colors should be warm, with traces left by real life. The characters should also feel real. We also added a small design at the opening, letting the cute handwritten characters of the title flash briefly on screen, like a life diary edited by oneself. Autoplay

The final generated film indeed has a good overall atmosphere. Kling was able to understand the old-neighborhood feel we wanted, and also arranged different shots such as the character walking, street details, and the indoor bookstore. Compared with the water-sleeve drum dance, which required precise contact relationships, this kind of short film focused on observing daily life clearly gave the model more room to perform.

But watching the film, there are still a few places that break the immersion. In the previous shot, the girl has just turned her head, but in the next shot she suddenly looks to the other side, and the action doesn't connect naturally. At the bookstore entrance, she reaches out and pushes open one door, and another door beside it unexpectedly opens by itself as well. The text on the bookstore sign is wrong, and although the street has old buildings, small shops, and plants, it is too clean and orderly, lacking the pedestrians and clutter commonly seen in real neighborhoods.

Looking at each shot individually, some of the images are indeed quite beautiful. But strung together into a 30-second short film, it is no longer just a test of whether the model can generate a few good-looking moving images. How a character moves from one shot into the next, and what interactions occur between the person and the environment—these details all affect whether the entire video looks natural.

Making a video, and making the video you want

Over the past few years, we have heard "everyone can become a director" far too many times when it comes to AI video. After seeing the sample films released by various companies, this statement seems increasingly convincing. Images that once required actors, cinematographers, and post-production teams to complete together can now potentially be generated by one person sitting at a computer.

But after actually using it, we have become somewhat hesitant about this statement.

In this test, more than once we wanted to take these decisions back into our own hands. We wanted the dancer to strike the drum with the tip of her sleeve, but the model made the drum move by itself; we wanted to capture the moment a girl naturally turns her head while walking, but the two consecutive shots did not connect.

Kling 4.0 certainly has its surprises. The rainy night in "Singin' in the Rain" has a strong cinematic feel, and video editing also successfully preserved the performance from the original footage. But the fact that seven or eight rounds of the water-sleeve dance still failed to strike the drum correctly, along with the credit consumption caused by repeated generation and editing, also made us realize that having a generation tool and being able to use it freely are still two different things.

Today, both Kling and Seedance are adding reference materials, extending video duration, and strengthening editing capabilities. For ordinary users, this means an opportunity to complete images that were previously difficult to produce independently; for teams that actually need to deliver content such as advertisements and short dramas, it means another question: whether they can stably obtain the desired results within an acceptable time and cost.

Generating video is becoming easier and easier. But once everyone can produce beautiful images, the tools that help people take fewer detours and require fewer rounds of rework may be the ones more likely to remain in use over the long term.